A tiny release between crawls (#138)

Bringing online new ranking changes

Co-authored-by: Viktor Lofgren <vlofgren@marginalia.nu>
Reviewed-on: https://git.marginalia.nu/marginalia/marginalia.nu/pulls/138

marginalia_nu/build.gradle

@@ -106,8 +106,9 @@ dependencies {
 
     implementation group: 'org.yaml', name: 'snakeyaml', version: '1.30'
 
-    implementation 'com.syncthemall:boilerpipe:1.2.2'
     implementation 'com.github.luben:zstd-jni:1.5.2-2'
+    implementation 'org.lz4:lz4-java:1.8.0'
+
     implementation 'com.github.vladimir-bukhtoyarov:bucket4j-core:7.5.0'
     implementation 'de.rototor.jeuclid:jeuclid-core:3.1.14'
 
@@ -126,7 +127,6 @@ dependencies {
     implementation 'org.roaringbitmap:RoaringBitmap:0.9.32'
 
     implementation group: 'mysql', name: 'mysql-connector-java', version: '8.0.29'
-
     implementation 'com.github.Marcono1234:gson-record-type-adapter-factory:0.2.0'
 
     testImplementation 'org.junit.jupiter:junit-jupiter-api:5.8.2'

marginalia_nu/src/main/java/nu/marginalia/util/BrailleBlockPunchCards.java

@@ -2,11 +2,13 @@ package nu.marginalia.util;
 
 public class BrailleBlockPunchCards {
 
+    private static final char brailleBlockBase = '\u2800';
+
     public static String printBits(int val, int bits) {
         StringBuilder builder = new StringBuilder();
 
         for (int b = 0; b < bits; b+=8, val>>>=8) {
-            builder.append((char)('\u2800'+bin2brail(val)));
+            builder.append((char)(brailleBlockBase + bin2brail(val)));
         }
 
         return builder.toString();

marginalia_nu/src/main/java/nu/marginalia/util/ParallelPipe.java

@@ -42,7 +42,7 @@ public abstract class ParallelPipe<INPUT,INTERMEDIATE> {
     @SneakyThrows
     private void runProcessThread() {
         while (expectingInput || !inputs.isEmpty()) {
-            var in = inputs.poll(1, TimeUnit.SECONDS);
+            var in = inputs.poll(10, TimeUnit.SECONDS);
 
             if (in != null) {
                 try {

marginalia_nu/src/main/java/nu/marginalia/util/RandomWriteFunnel.java

@@ -108,7 +108,6 @@ public class RandomWriteFunnel implements AutoCloseable {
 
         private void eval(ByteBuffer dest) throws IOException {
             flushBuffer();
-            channel.force(false);
 
             channel.position(0);
             buffer.clear();

marginalia_nu/src/main/java/nu/marginalia/util/StringPool.java

@@ -1,20 +1,33 @@
 package nu.marginalia.util;
 
+import it.unimi.dsi.fastutil.objects.Object2LongOpenHashMap;
+
+import java.util.Arrays;
 import java.util.HashMap;
 
 public class StringPool {
-    private final HashMap<String, String> words;
 
-    public StringPool() {
-        this.words  = new HashMap<>(1000);
+    private final HashMap<String, String> words;
+    private final Object2LongOpenHashMap<String> ages;
+    private final int maxCap;
+
+    long idx;
+
+    private StringPool(int capacity, int maxCap) {
+        this.ages  = new Object2LongOpenHashMap<>(capacity);
+        this.words = new HashMap<>(capacity);
+        this.maxCap = maxCap;
     }
 
-    public StringPool(int capacity) {
-        words = new HashMap<>(capacity);
+    public static StringPool create(int capacity) {
+        return new StringPool(capacity, capacity * 10);
     }
 
     public String internalize(String str) {
+        prune();
+
         final String ret = words.putIfAbsent(str, str);
+        ages.put(ret, idx++);
 
         if (null == ret)
             return str;
@@ -22,6 +35,37 @@ public class StringPool {
         return ret;
     }
 
+    public String[] internalize(String[] str) {
+
+        for (int i = 0; i < str.length; i++) {
+            str[i] = internalize(str[i]);
+        }
+
+        return str;
+    }
+
+    public void prune() {
+
+        if (words.size() < maxCap)
+            return;
+
+        long[] ageValues = ages.values().toLongArray();
+        Arrays.sort(ageValues);
+
+        long cutoff = ageValues[ageValues.length - maxCap / 10];
+
+        words.clear();
+        ages.forEach((word, cnt) -> {
+            if (cnt >= cutoff) {
+                words.put(word, word);
+            }
+        });
+        ages.clear();
+        words.forEach((w,w2) -> {
+            ages.put(w, idx);
+        });
+    }
+
     public void flush() {
         words.clear();
     }

marginalia_nu/src/main/java/nu/marginalia/util/array/algo/IntArrayBase.java

@@ -18,22 +18,14 @@ public interface IntArrayBase extends BulkTransferArray<IntBuffer> {
         }
     }
 
-    default void increment(long pos) {
-        set(pos, get(pos) + 1);
-    }
-
     default void swap(long pos1, long pos2) {
         int tmp = get(pos1);
         set(pos1, get(pos2));
         set(pos2, tmp);
     }
 
-    default void swapn(int n, long pos1, long pos2) {
-        for (int i = 0; i < n; i++) {
-            int tmp = get(pos1+i);
-            set(pos1+i, get(pos2+i));
-            set(pos2+i, tmp);
-        }
+    default void increment(long pos) {
+        set(pos, get(pos) + 1);
     }
 
     default int getAndIncrement(long pos) {
@@ -47,6 +39,7 @@ public interface IntArrayBase extends BulkTransferArray<IntBuffer> {
             set(start+i, buffer.get(i + bufferStart));
         }
     }
+
     default void get(long start, long end, IntBuffer buffer, int bufferStart) {
         for (int i = 0; i < (end-start); i++) {
             buffer.put(i + bufferStart, get(start + i));

marginalia_nu/src/main/java/nu/marginalia/util/array/algo/LongArrayBase.java

@@ -28,6 +28,7 @@ public interface LongArrayBase extends BulkTransferArray<LongBuffer> {
         set(pos2, tmp);
     }
 
+    /** Behavior not defined for overlapping ranges */
     default void swapn(int n, long pos1, long pos2) {
         for (int i = 0; i < n; i++) {
             long tmp = get(pos1+i);

marginalia_nu/src/main/java/nu/marginalia/util/array/delegate/ShiftedIntArray.java

@@ -2,6 +2,7 @@ package nu.marginalia.util.array.delegate;
 
 import com.upserve.uppend.blobs.NativeIO;
 import nu.marginalia.util.array.IntArray;
+import nu.marginalia.util.array.algo.SortingContext;
 import nu.marginalia.util.array.buffer.IntQueryBuffer;
 import nu.marginalia.util.array.functional.IntBinaryIOOperation;
 import nu.marginalia.util.array.functional.IntIOTransformer;
@@ -61,6 +62,16 @@ public class ShiftedIntArray implements IntArray {
         delegate.get(shift+start, shift+end, buffer);
     }
 
+    @Override
+    public int getAndIncrement(long pos) {
+        return delegate.getAndIncrement(shift + pos);
+    }
+
+    @Override
+    public void fill(long start, long end, int val) {
+        delegate.fill(start + shift, end + shift, val);
+    }
+
     @Override
     public long size() {
         return size;
@@ -97,6 +108,12 @@ public class ShiftedIntArray implements IntArray {
         return delegate.isSorted(shift + start, shift + end);
     }
 
+
+    public void sortLargeSpan(SortingContext ctx, long start, long end) throws IOException {
+        delegate.sortLargeSpan(ctx, start, end);
+    }
+
+
     public long search(int key) {
         if (size < 128) {
             return linearSearch(key);

marginalia_nu/src/main/java/nu/marginalia/util/array/delegate/ShiftedLongArray.java

@@ -3,6 +3,7 @@ package nu.marginalia.util.array.delegate;
 import com.upserve.uppend.blobs.NativeIO;
 import nu.marginalia.util.array.LongArray;
 import nu.marginalia.util.array.algo.LongArraySearch;
+import nu.marginalia.util.array.algo.SortingContext;
 import nu.marginalia.util.array.buffer.LongQueryBuffer;
 import nu.marginalia.util.array.functional.LongBinaryIOOperation;
 import nu.marginalia.util.array.functional.LongIOTransformer;
@@ -62,6 +63,16 @@ public class ShiftedLongArray implements LongArray {
         delegate.get(shift+start, shift+end, buffer);
     }
 
+    @Override
+    public long getAndIncrement(long pos) {
+        return delegate.getAndIncrement(shift + pos);
+    }
+
+    @Override
+    public void fill(long start, long end, long val) {
+        delegate.fill(start + shift, end + shift, val);
+    }
+
     @Override
     public long size() {
         return size;
@@ -106,6 +117,14 @@ public class ShiftedLongArray implements LongArray {
         return delegate.isSortedN(sz, shift + start, shift + end);
     }
 
+    public void sortLargeSpanN(SortingContext ctx, int sz, long start, long end) throws IOException {
+        delegate.sortLargeSpanN(ctx, sz, start, end);
+    }
+
+    public void sortLargeSpan(SortingContext ctx, long start, long end) throws IOException {
+        delegate.sortLargeSpan(ctx, start, end);
+    }
+
     public long searchN(int sz, long key) {
         if (size < 128) {
             return linearSearchN(sz, key);

marginalia_nu/src/main/java/nu/marginalia/util/array/page/PagingIntArray.java

@@ -2,6 +2,7 @@ package nu.marginalia.util.array.page;
 
 import com.upserve.uppend.blobs.NativeIO;
 import nu.marginalia.util.array.IntArray;
+import nu.marginalia.util.array.algo.SortingContext;
 import nu.marginalia.util.array.buffer.IntQueryBuffer;
 import nu.marginalia.util.array.delegate.ReferenceImplIntArrayDelegate;
 import nu.marginalia.util.array.functional.IntBinaryIOOperation;
@@ -113,6 +114,11 @@ public class PagingIntArray extends AbstractPagingArray<IntArrayPage, IntBuffer>
         }
     }
 
+    @Override
+    public int getAndIncrement(long pos) {
+        return pages[partitioningScheme.getPage(pos)].getAndIncrement(partitioningScheme.getOffset(pos));
+    }
+
     @Override
     public void get(long start, long end, int[] buffer) {
         if (partitioningScheme.isSamePage(start, end)) {
@@ -272,6 +278,22 @@ public class PagingIntArray extends AbstractPagingArray<IntArrayPage, IntBuffer>
         }
     }
 
+
+    public void sortLargeSpan(SortingContext ctx, long start, long end) throws IOException {
+        if (partitioningScheme.isSamePage(start, end)) {
+            int sOff = partitioningScheme.getOffset(start);
+            int eOff = partitioningScheme.getEndOffset(start, end);
+
+            if (eOff > sOff) {
+                pages[partitioningScheme.getPage(start)].sortLargeSpan(ctx, sOff, eOff);
+            }
+        }
+        else {
+            defaults.sortLargeSpan(ctx, start, end);
+        }
+    }
+
+
     public void write(Path fileName) throws IOException {
         try (var channel = (FileChannel) Files.newByteChannel(fileName, StandardOpenOption.CREATE, StandardOpenOption.WRITE)) {
             for (int i = 0; i < pages.length; i++) {

marginalia_nu/src/main/java/nu/marginalia/util/array/page/PagingLongArray.java

@@ -2,6 +2,7 @@ package nu.marginalia.util.array.page;
 
 import com.upserve.uppend.blobs.NativeIO;
 import nu.marginalia.util.array.LongArray;
+import nu.marginalia.util.array.algo.SortingContext;
 import nu.marginalia.util.array.buffer.LongQueryBuffer;
 import nu.marginalia.util.array.delegate.ReferenceImplLongArrayDelegate;
 import nu.marginalia.util.array.functional.LongBinaryIOOperation;
@@ -118,6 +119,11 @@ public class PagingLongArray extends AbstractPagingArray<LongArrayPage, LongBuff
         }
     }
 
+    @Override
+    public long getAndIncrement(long pos) {
+        return pages[partitioningScheme.getPage(pos)].getAndIncrement(partitioningScheme.getOffset(pos));
+    }
+
     @Override
     public void set(long pos, long value) {
         int page = partitioningScheme.getPage(pos);
@@ -439,6 +445,33 @@ public class PagingLongArray extends AbstractPagingArray<LongArrayPage, LongBuff
             defaults.mergeSortN(sz, start, end, tempDir);
         }
     }
+    public void sortLargeSpanN(SortingContext ctx, int sz, long start, long end) throws IOException {
+        if (partitioningScheme.isSamePage(start, end)) {
+            int sOff = partitioningScheme.getOffset(start);
+            int eOff = partitioningScheme.getEndOffset(start, end);
+
+            if (eOff > sOff) {
+                pages[partitioningScheme.getPage(start)].sortLargeSpanN(ctx, sz, sOff, eOff);
+            }
+        }
+        else {
+            defaults.sortLargeSpanN(ctx, sz, start, end);
+        }
+    }
+
+    public void sortLargeSpan(SortingContext ctx, long start, long end) throws IOException {
+        if (partitioningScheme.isSamePage(start, end)) {
+            int sOff = partitioningScheme.getOffset(start);
+            int eOff = partitioningScheme.getEndOffset(start, end);
+
+            if (eOff > sOff) {
+                pages[partitioningScheme.getPage(start)].sortLargeSpan(ctx, sOff, eOff);
+            }
+        }
+        else {
+            defaults.sortLargeSpan(ctx, start, end);
+        }
+    }
 
     public void write(Path fileName) throws IOException {
         try (var channel = (FileChannel) Files.newByteChannel(fileName, StandardOpenOption.CREATE, StandardOpenOption.WRITE)) {

marginalia_nu/src/main/java/nu/marginalia/util/bigstring/BigString.java

@@ -0,0 +1,17 @@
+package nu.marginalia.util.bigstring;
+
+public interface BigString {
+    static BigString encode(String stringValue) {
+        if (stringValue.length() > 64) {
+            return new CompressedBigString(stringValue);
+        }
+        else {
+            return new PlainBigString(stringValue);
+        }
+    }
+    String decode();
+
+    byte[] getBytes();
+
+    int length();
+}

marginalia_nu/src/main/java/nu/marginalia/util/bigstring/CompressedBigString.java

@@ -0,0 +1,39 @@
+package nu.marginalia.util.bigstring;
+
+import net.jpountz.lz4.LZ4Compressor;
+import net.jpountz.lz4.LZ4Factory;
+import net.jpountz.lz4.LZ4FastDecompressor;
+
+import java.nio.charset.StandardCharsets;
+
+public class CompressedBigString implements BigString {
+    private final int originalSize;
+    private final int length;
+    private final byte[] encoded;
+
+    private static final LZ4Factory lz4Factory = LZ4Factory.fastestInstance();;
+    private static final LZ4Compressor compressor = lz4Factory.fastCompressor();
+    private static final LZ4FastDecompressor decompressor = lz4Factory.fastDecompressor();
+
+    public CompressedBigString(String stringValue) {
+        byte[] byteValue = stringValue.getBytes(StandardCharsets.UTF_16);
+        originalSize = byteValue.length;
+        encoded = compressor.compress(byteValue);
+        length = stringValue.length();
+    }
+
+    @Override
+    public String decode() {
+        return new String(getBytes(), StandardCharsets.UTF_16);
+    }
+
+    @Override
+    public byte[] getBytes() {
+        return decompressor.decompress(encoded, originalSize);
+    }
+
+    @Override
+    public int length() {
+        return length;
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/util/bigstring/PlainBigString.java

@@ -0,0 +1,26 @@
+package nu.marginalia.util.bigstring;
+
+import java.nio.charset.StandardCharsets;
+
+public class PlainBigString implements BigString {
+    private final String value;
+
+    public PlainBigString(String value) {
+        this.value = value;
+    }
+
+    @Override
+    public String decode() {
+        return value;
+    }
+
+    @Override
+    public byte[] getBytes() {
+        return value.getBytes(StandardCharsets.UTF_8);
+    }
+
+    @Override
+    public int length() {
+        return value.length();
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/util/dict/DictionaryData.java

@@ -1,8 +1,5 @@
 package nu.marginalia.util.dict;
 
-import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
-
 import java.nio.ByteBuffer;
 import java.nio.LongBuffer;
 import java.util.ArrayList;
@@ -10,7 +7,6 @@ import java.util.ArrayList;
 public class DictionaryData {
 
     private final int DICTIONARY_BANK_SIZE;
-    private static final Logger logger = LoggerFactory.getLogger(DictionaryData.class);
 
     private final ArrayList<DictionaryDataBank> banks = new ArrayList<>(100);
 

marginalia_nu/src/main/java/nu/marginalia/util/dict/DictionaryMap.java

@@ -1,6 +1,17 @@
 package nu.marginalia.util.dict;
 
 public interface DictionaryMap {
+    int NO_VALUE = Integer.MIN_VALUE;
+
+    static DictionaryMap create() {
+        if (Boolean.getBoolean("small-ram")) {
+            return new OnHeapDictionaryMap();
+        }
+        else {
+            return new OffHeapDictionaryHashMap(1L << 31);
+        }
+    }
+
     int size();
 
     int put(long key);

marginalia_nu/src/main/java/nu/marginalia/util/dict/OffHeapDictionaryHashMap.java

@@ -16,15 +16,14 @@ import static nu.marginalia.util.FileSizeUtil.readableSize;
  *  Spiritually influenced by GNU Trove's hash maps
  *  LGPL 2.1
  */
-public class DictionaryHashMap implements DictionaryMap {
-    private static final Logger logger = LoggerFactory.getLogger(DictionaryHashMap.class);
+public class OffHeapDictionaryHashMap implements DictionaryMap {
+    private static final Logger logger = LoggerFactory.getLogger(OffHeapDictionaryHashMap.class);
     private static final Gauge probe_count_metrics
             = Gauge.build("wmsa_dictionary_hash_map_probe_count", "Probing Count")
             .register();
 
     private final int bufferCount;
     private final IntBuffer[] buffers;
-    public static final int NO_VALUE = Integer.MIN_VALUE;
 
     private final DictionaryData dictionaryData;
 
@@ -35,7 +34,7 @@ public class DictionaryHashMap implements DictionaryMap {
 
     private final AtomicInteger sz = new AtomicInteger(0);
 
-    public DictionaryHashMap(long sizeMemory) {
+    public OffHeapDictionaryHashMap(long sizeMemory) {
         final int intSize = 4;
 
         bufferCount = 1 + (int) ((intSize*sizeMemory) / (1<<30));

marginalia_nu/src/main/java/nu/marginalia/util/dict/OnHeapDictionaryMap.java

@@ -0,0 +1,23 @@
+package nu.marginalia.util.dict;
+
+import it.unimi.dsi.fastutil.longs.Long2IntOpenHashMap;
+
+public class OnHeapDictionaryMap implements DictionaryMap {
+    private final Long2IntOpenHashMap entries = new Long2IntOpenHashMap(100_000, 0.75f);
+
+    @Override
+    public int size() {
+        return entries.size();
+    }
+
+    @Override
+    public int put(long key) {
+        entries.putIfAbsent(key, entries.size());
+        return get(key);
+    }
+
+    @Override
+    public int get(long key) {
+        return entries.getOrDefault(key, NO_VALUE);
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/util/gregex/GuardedRegexFactory.java

@@ -19,9 +19,6 @@ public class GuardedRegexFactory {
     public static GuardedRegex contains(String substring, @Language("RegExp") String regex) {
         return new GuardedRegexContains(substring, regex);
     }
-    public static GuardedRegex minLength(int minLength, @Language("RegExp") String regex) {
-        return new GuardedRegexMinLength(minLength, regex);
-    }
 
     private record GuardedRegexContains(String contains, Pattern pattern) implements GuardedRegex {
         public GuardedRegexContains(String contains, String pattern) {
@@ -32,15 +29,6 @@ public class GuardedRegexFactory {
             return s.contains(contains) && pattern.matcher(s).find();
         }
     }
-    private record GuardedRegexMinLength(int minLength, Pattern pattern) implements GuardedRegex {
-        public GuardedRegexMinLength(int minLength, String pattern) {
-            this(minLength, Pattern.compile(pattern));
-        }
-
-        public boolean test(String s) {
-            return s.length() >= minLength && pattern.matcher(s).find();
-        }
-    }
     private record GuardedRegexStartsWith(String start,  Pattern pattern) implements GuardedRegex {
         public GuardedRegexStartsWith(String start, String pattern) {
             this(start, Pattern.compile(pattern));

marginalia_nu/src/main/java/nu/marginalia/util/language/DocumentDebugger.java

@@ -4,7 +4,7 @@ import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.KeywordCounter;
 import nu.marginalia.util.language.processing.KeywordExtractor;
 import nu.marginalia.util.language.processing.NameCounter;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordRep;
 import nu.marginalia.util.language.processing.model.tag.WordSeparator;
@@ -68,9 +68,6 @@ public class DocumentDebugger {
 
         Set<String> reps = new HashSet<>();
 
-//        kc.count(languageData, 0.75).forEach(rep -> reps.add(rep.stemmed));
-//        kc.count(languageData).forEach(rep -> reps.add(rep.stemmed));
-
         try (var pw = new PrintWriter(new FileOutputStream(output.toFile()))) {
 
             for (var sent : languageData.titleSentences) {

marginalia_nu/src/main/java/nu/marginalia/util/language/WordPatterns.java

@@ -1,5 +1,7 @@
 package nu.marginalia.util.language;
 
+import org.apache.commons.lang3.StringUtils;
+
 import java.io.BufferedReader;
 import java.io.IOException;
 import java.io.InputStreamReader;
@@ -123,14 +125,24 @@ public class WordPatterns {
         if (!filter(s)) {
             return true;
         }
-        if (isTopWord(s)) {
+
+        String sLc;
+        if (StringUtils.isAllLowerCase(s)) {
+            sLc = s;
+        }
+        else {
+            sLc = s.toLowerCase();
+        }
+
+        if (isTopWord(sLc)) {
             return true;
         }
+
         return false;
     }
 
-    public static boolean isTopWord(String s) {
-        return topWords.contains(s.toLowerCase());
+    public static boolean isTopWord(String strLowerCase) {
+        return topWords.contains(strLowerCase);
     }
 
 }

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/DocumentKeywordExtractor.java

@@ -35,9 +35,7 @@ public class DocumentKeywordExtractor {
 
         List<WordRep> titleWords = extractTitleWords(documentLanguageData);
         List<WordRep> wordsNamesAll = nameCounter.count(documentLanguageData, 2);
-        List<WordRep> subjects = subjectCounter.count(documentLanguageData);
-
-        tfIdfCounter.countHisto(keywordMetadata, documentLanguageData);
+        List<WordRep> subjects = subjectCounter.count(keywordMetadata, documentLanguageData);
 
         for (var rep : titleWords) keywordMetadata.titleKeywords().add(rep.stemmed);
         for (var rep : wordsNamesAll) keywordMetadata.namesKeywords().add(rep.stemmed);
@@ -59,11 +57,12 @@ public class DocumentKeywordExtractor {
 
         getWordPositions(keywordMetadata, documentLanguageData);
 
-        List<WordRep> wordsNamesAll = nameCounter.count(documentLanguageData, 2);
-        List<WordRep> subjects = subjectCounter.count(documentLanguageData);
-
         List<WordRep> wordsTfIdf = tfIdfCounter.countHisto(keywordMetadata, documentLanguageData);
 
+        List<WordRep> wordsNamesAll = nameCounter.count(documentLanguageData, 2);
+        List<WordRep> subjects = subjectCounter.count(keywordMetadata, documentLanguageData);
+
+
         for (var rep : titleWords) keywordMetadata.titleKeywords().add(rep.stemmed);
         for (var rep : wordsNamesAll) keywordMetadata.namesKeywords().add(rep.stemmed);
         for (var rep : subjects) keywordMetadata.subjectKeywords().add(rep.stemmed);
@@ -94,7 +93,7 @@ public class DocumentKeywordExtractor {
                 ret.merge(word.stemmed(), posBit, this::bitwiseOr);
             }
 
-            for (var span : keywordExtractor.getNames(sent)) {
+            for (var span : keywordExtractor.getProperNames(sent)) {
                 ret.merge(sent.constructStemmedWordFromSpan(span), posBit, this::bitwiseOr);
             }
         }
@@ -108,7 +107,7 @@ public class DocumentKeywordExtractor {
                 ret.merge(word.stemmed(), posBit, this::bitwiseOr);
             }
 
-            for (var span : keywordExtractor.getNames(sent)) {
+            for (var span : keywordExtractor.getProperNames(sent)) {
                 ret.merge(sent.constructStemmedWordFromSpan(span), posBit, this::bitwiseOr);
             }
 
@@ -155,16 +154,16 @@ public class DocumentKeywordExtractor {
                 if (!word.isStopWord()) {
                     String w = AsciiFlattener.flattenUnicode(word.wordLowerCase());
                     if (WordPatterns.singleWordQualitiesPredicate.test(w)) {
-                        wordsBuilder.add(w, metadata.forWord(flagsTemplate, word.stemmed()));
+                        wordsBuilder.add(w, metadata.getMetadataForWord(flagsTemplate, word.stemmed()));
                     }
                 }
             }
 
-            for (var names : keywordExtractor.getNames(sent)) {
+            for (var names : keywordExtractor.getProperNames(sent)) {
                 var rep = new WordRep(sent, names);
                 String w = AsciiFlattener.flattenUnicode(rep.word);
 
-                wordsBuilder.add(w, metadata.forWord(flagsTemplate, rep.stemmed));
+                wordsBuilder.add(w, metadata.getMetadataForWord(flagsTemplate, rep.stemmed));
             }
         }
 
@@ -218,7 +217,7 @@ public class DocumentKeywordExtractor {
                 continue;
             }
 
-            wordsBuilder.add(flatWord, metadata.forWord(metadata.wordFlagsTemplate(), word.stemmed) | additionalMeta);
+            wordsBuilder.add(flatWord, metadata.getMetadataForWord(metadata.wordFlagsTemplate(), word.stemmed) | additionalMeta);
         }
     }
 

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/KeywordCounter.java

@@ -43,8 +43,8 @@ public class KeywordCounter {
 
                 counts.mergeInt(rep.stemmed, 1, Integer::sum);
 
-                var instanceSet = instances.computeIfAbsent(rep.stemmed, k -> new HashSet<>(500));
-                if (instanceSet.size() < 250) {
+                var instanceSet = instances.computeIfAbsent(rep.stemmed, k -> new HashSet<>(16));
+                if (instanceSet.size() < 4) {
                     instanceSet.add(rep);
                 }
             }

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/KeywordExtractor.java

@@ -7,14 +7,12 @@ import nu.marginalia.util.language.processing.model.tag.WordSeparator;
 
 import java.lang.ref.SoftReference;
 import java.util.ArrayList;
-import java.util.Collections;
 import java.util.List;
-import java.util.Set;
 
 public class KeywordExtractor {
 
-    public WordSpan[] getNames(DocumentSentence sentence) {
-        List<WordSpan> spans = new ArrayList<>(sentence.length());
+    public WordSpan[] getProperNames(DocumentSentence sentence) {
+        List<WordSpan> spans = new ArrayList<>(2 * sentence.length());
 
         for (int i = 0; i < sentence.length(); i++) {
             if (isProperNoun(i, sentence))
@@ -57,27 +55,73 @@ public class KeywordExtractor {
         return spans.toArray(WordSpan[]::new);
     }
 
-    public WordSpan[] getKeywordsFromSentence(DocumentSentence sentence) {
-        if (sentence.keywords != null) {
-            return sentence.keywords.get();
-        }
-        List<WordSpan> spans = new ArrayList<>(sentence.length());
 
-        Set<String> topWords = Collections.emptySet();
+    public WordSpan[] getNouns(DocumentSentence sentence) {
+        List<WordSpan> spans = new ArrayList<>(2 * sentence.length());
 
         for (int i = 0; i < sentence.length(); i++) {
-            if (isName(i, sentence, topWords) || isTopAdj(i, sentence, topWords))
+            if (isNoun(i, sentence))
                 spans.add(new WordSpan(i, i+1));
         }
 
         for (int i = 1; i < sentence.length(); i++) {
             if (sentence.separators[i-1] == WordSeparator.COMMA) { continue; }
 
-            if (isName(i, sentence, topWords)) {
-                if (isName(i - 1, sentence, topWords) || isTopAdj(i-1, sentence, topWords))
+            if (isNoun(i, sentence)
+                    && (isNoun(i-1, sentence)) || "JJ".equals(sentence.posTags[i-1])) {
+                spans.add(new WordSpan(i - 1, i + 1));
+            }
+        }
+
+        for (int i = 2; i < sentence.length(); i++) {
+            if (sentence.separators[i-2] == WordSeparator.COMMA) { continue; }
+            if (sentence.separators[i-1] == WordSeparator.COMMA) { i++; continue; }
+
+            if ((isNoun(i, sentence))
+                    && (isJoiner(sentence, i-1) || isNoun(i-1, sentence))
+                    && (isNoun(i-2, sentence)) || "JJ".equals(sentence.posTags[i-2]))
+                spans.add(new WordSpan(i-2, i+1));
+        }
+
+        for (int i = 3; i < sentence.length(); i++) {
+            if (sentence.separators[i-3] == WordSeparator.COMMA) { continue; }
+            if (sentence.separators[i-2] == WordSeparator.COMMA) { i++; continue; }
+            if (sentence.separators[i-1] == WordSeparator.COMMA) { i+=2; continue; }
+
+            if (isNoun(i, sentence) && (isNoun(i-3, sentence) || "JJ".equals(sentence.posTags[i-3]))) {
+                if (isNoun(i - 1, sentence) && isNoun(i - 2, sentence))
+                    spans.add(new WordSpan(i-3, i+1));
+                else if (isJoiner(sentence, i-2) && sentence.posTags[i-1].equals("DT"))
+                    spans.add(new WordSpan(i-3, i+1));
+                else if ((isJoiner(sentence, i-1) ||isNoun(i-1, sentence))
+                        && (isJoiner(sentence, i-2)||isNoun(i-2, sentence)))
+                    spans.add(new WordSpan(i-3, i+1));
+            }
+        }
+
+        return spans.toArray(WordSpan[]::new);
+    }
+
+
+    public WordSpan[] getKeywordsFromSentence(DocumentSentence sentence) {
+        if (sentence.keywords != null) {
+            return sentence.keywords.get();
+        }
+        List<WordSpan> spans = new ArrayList<>(2 * sentence.length());
+
+        for (int i = 0; i < sentence.length(); i++) {
+            if (isName(i, sentence) || isTopAdj(i, sentence))
+                spans.add(new WordSpan(i, i+1));
+        }
+
+        for (int i = 1; i < sentence.length(); i++) {
+            if (sentence.separators[i-1] == WordSeparator.COMMA) { continue; }
+
+            if (isName(i, sentence)) {
+                if (isName(i - 1, sentence) || isTopAdj(i-1, sentence))
                     spans.add(new WordSpan(i - 1, i + 1));
             }
-            if (sentence.posTags[i].equals("CD") &&  isName(i-1, sentence, topWords)) {
+            if (sentence.posTags[i].equals("CD") &&  isName(i-1, sentence)) {
                 spans.add(new WordSpan(i - 1, i + 1));
             }
         }
@@ -86,16 +130,16 @@ public class KeywordExtractor {
             if (sentence.separators[i-1] == WordSeparator.COMMA) { i++; continue; }
             if (sentence.separators[i-2] == WordSeparator.COMMA) { continue; }
 
-            if (isName(i, sentence, topWords)) {
-                if ((isName(i-1, sentence, topWords) || isTopAdj(i-1, sentence, topWords))
-                        && (isName(i-2, sentence, topWords) || isTopAdj(i-2, sentence, topWords))) {
+            if (isName(i, sentence)) {
+                if ((isName(i-1, sentence) || isTopAdj(i-1, sentence))
+                        && (isName(i-2, sentence) || isTopAdj(i-2, sentence))) {
                     spans.add(new WordSpan(i - 2, i + 1));
                 }
                 else if ((isProperNoun(i-1, sentence) || isJoiner(sentence, i-1)) && isProperNoun(i-2, sentence)) {
                     spans.add(new WordSpan(i - 2, i + 1));
                 }
             }
-            else if (sentence.posTags[i].equals("CD") && isName(i-1, sentence, topWords) && isName(i-2, sentence, topWords)) {
+            else if (sentence.posTags[i].equals("CD") && isName(i-1, sentence) && isName(i-2, sentence)) {
                 spans.add(new WordSpan(i - 2, i + 1));
             }
         }
@@ -105,10 +149,10 @@ public class KeywordExtractor {
             if (sentence.separators[i-2] == WordSeparator.COMMA) { i++; continue; }
             if (sentence.separators[i-3] == WordSeparator.COMMA) { continue; }
 
-            if (isName(i, sentence, topWords) &&
-                    (isName(i-1, sentence, topWords) || isTopAdj(i-1, sentence, topWords)) &&
-                    (isName(i-2, sentence, topWords) || isTopAdj(i-2, sentence, topWords)) &&
-                    (isName(i-3, sentence, topWords) || isTopAdj(i-3, sentence, topWords))) {
+            if (isName(i, sentence) &&
+                    (isName(i-1, sentence) || isTopAdj(i-1, sentence)) &&
+                    (isName(i-2, sentence) || isTopAdj(i-2, sentence)) &&
+                    (isName(i-3, sentence) || isTopAdj(i-3, sentence))) {
                 spans.add(new WordSpan(i - 3, i + 1));
             }
             else if (isProperNoun(i, sentence) && isProperNoun(i-3, sentence)) {
@@ -134,7 +178,9 @@ public class KeywordExtractor {
     public boolean isProperNoun(int i, DocumentSentence sent) {
         return "NNP".equals(sent.posTags[i]) || "NNPS".equals(sent.posTags[i]);
     }
-
+    public boolean isNoun(int i, DocumentSentence sent) {
+        return sent.posTags[i].startsWith("NN");
+    }
     public boolean isJoiner(DocumentSentence sent, int i) {
         if(sent.posTags[i].equals("IN")) {
             return true;
@@ -183,21 +229,13 @@ public class KeywordExtractor {
         return true;
     }
 
-    private boolean isName(int i, DocumentSentence sentence, Set<String> topWords) {
-        if (!topWords.isEmpty()) {
-            String posTag = sentence.posTags[i];
-            String word = sentence.stemmedWords[i];
-
-            return ((topWords.contains(word)) && (posTag.startsWith("N") || "VBN".equals(posTag)) && !sentence.isStopWord(i));
-        }
-
-
+    private boolean isName(int i, DocumentSentence sentence) {
         String posTag = sentence.posTags[i];
 
-        return (posTag.startsWith("N") || "VBN".equals(posTag)) && !sentence.isStopWord(i);
+        return (posTag.startsWith("N") || "VBG".equals(posTag)|| "VBN".equals(posTag)) && !sentence.isStopWord(i);
     }
 
-    private boolean isTopAdj(int i, DocumentSentence sentence, Set<String> topWords) {
+    private boolean isTopAdj(int i, DocumentSentence sentence) {
         String posTag = sentence.posTags[i];
 
         return (posTag.startsWith("JJ") || posTag.startsWith("R") || posTag.startsWith("VBG"));

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/NameCounter.java

@@ -20,7 +20,7 @@ public class NameCounter {
 
         for (int i = 0; i < dld.sentences.length; i++) {
             DocumentSentence sent = dld.sentences[i];
-            var keywords = keywordExtractor.getNames(sent);
+            var keywords = keywordExtractor.getProperNames(sent);
             for (var span : keywords) {
                 if (span.size() <= 1)
                     continue;

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/SubjectCounter.java

@@ -1,9 +1,11 @@
 package nu.marginalia.util.language.processing;
 
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
+import nu.marginalia.util.language.processing.model.KeywordMetadata;
 import nu.marginalia.util.language.processing.model.WordRep;
 import nu.marginalia.util.language.processing.model.WordSpan;
 import nu.marginalia.util.language.processing.model.tag.WordSeparator;
+import org.apache.commons.lang3.StringUtils;
 
 import java.util.*;
 import java.util.stream.Collectors;
@@ -23,13 +25,13 @@ public class SubjectCounter {
     // Greeks bearing gifts -> Greeks
     // Steve McQueen drove fast | cars -> Steve McQueen
 
-    public List<WordRep> count(DocumentLanguageData dld) {
+    public List<WordRep> count(KeywordMetadata keywordMetadata, DocumentLanguageData dld) {
 
         Map<String, Integer> counts = new HashMap<>();
         Map<String, Set<WordRep>> instances = new HashMap<>();
 
         for (var sentence : dld.sentences) {
-            for (WordSpan kw : keywordExtractor.getNames(sentence)) {
+            for (WordSpan kw : keywordExtractor.getNouns(sentence)) {
                 if (kw.end + 2 >= sentence.length()) {
                     continue;
                 }
@@ -46,20 +48,46 @@ public class SubjectCounter {
 
                     String stemmed = rep.stemmed;
 
-                    counts.merge(stemmed, -1, Integer::sum);
                     instances.computeIfAbsent(stemmed, s -> new HashSet<>()).add(rep);
                 }
             }
         }
 
-        int best = counts.values().stream().mapToInt(Integer::valueOf).min().orElse(0);
+        Map<String, Integer> scores = new HashMap<>(instances.size());
+        for (String stemmed : instances.keySet()) {
+            scores.put(stemmed, getTermTfIdf(keywordMetadata, stemmed));
+        }
 
-        return counts.entrySet().stream().sorted(Map.Entry.comparingByValue())
-                .filter(e -> e.getValue()<-2 && e.getValue()<=best*0.75)
+        return scores.entrySet().stream()
+                .filter(e -> e.getValue() >= 150)
                 .flatMap(e -> instances.getOrDefault(e.getKey(), Collections.emptySet()).stream())
                 .collect(Collectors.toList());
     }
 
+    private int getTermTfIdf(KeywordMetadata keywordMetadata, String stemmed) {
+        if (stemmed.contains("_")) {
+            int sum = 0;
+            String[] parts = StringUtils.split(stemmed, '_');
+
+            if (parts.length == 0) {
+                return  0;
+            }
+
+            for (String part : parts) {
+                sum += getTermTfIdf(keywordMetadata, part);
+            }
+
+            return sum / parts.length;
+        }
+
+        var meta = keywordMetadata.wordsTfIdf().get(stemmed);
+        if (meta != null) {
+            return meta.tfIdfNormalized();
+        }
+
+        return 0;
+    }
+
     private boolean isDetOrAdverbOrVerb(String posTag) {
         return "DT".equals(posTag) // determinant
                 || "RB".equals(posTag) // adverb

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/model/DocumentLanguageData.java

@@ -2,12 +2,13 @@ package nu.marginalia.util.language.processing.model;
 
 import gnu.trove.map.hash.TObjectIntHashMap;
 import lombok.AllArgsConstructor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 
 import java.util.Arrays;
 import java.util.stream.Stream;
 
 /**
- * @see nu.marginalia.util.language.processing.SentenceExtractor
+ * @see SentenceExtractor
  */
 @AllArgsConstructor
 public class DocumentLanguageData {

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/model/KeywordMetadata.java

@@ -17,9 +17,6 @@ public record KeywordMetadata(HashSet<String> titleKeywords,
 )
 {
 
-    private static final KeywordCounter.WordFrequencyData empty = new KeywordCounter.WordFrequencyData(0, 0);
-    private static final int TF_IDF_HIGH_LIMIT = 64;
-
     public KeywordMetadata(EnumSet<EdgePageWordFlags> flags) {
         this(new HashSet<>(50), new HashSet<>(10), new HashSet<>(50),
                 new HashMap<>(15_000),
@@ -31,7 +28,8 @@ public record KeywordMetadata(HashSet<String> titleKeywords,
         this(EnumSet.noneOf(EdgePageWordFlags.class));
     }
 
-    public long forWord(EnumSet<EdgePageWordFlags> flagsTemplate, String stemmed) {
+    private static final KeywordCounter.WordFrequencyData empty = new KeywordCounter.WordFrequencyData(0, 0);
+    public long getMetadataForWord(EnumSet<EdgePageWordFlags> flagsTemplate, String stemmed) {
 
         KeywordCounter.WordFrequencyData tfidf = wordsTfIdf.getOrDefault(stemmed, empty);
         EnumSet<EdgePageWordFlags> flags = flagsTemplate.clone();

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/model/WordRep.java

@@ -24,7 +24,7 @@ public class WordRep implements Comparable<WordRep> {
 
     @Override
     public int compareTo(@NotNull WordRep o) {
-        return stemmed.compareTo(o.stemmed);
+        return word.compareTo(o.word);
     }
 
     @Override

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/sentence/SentenceExtractor.java

@@ -1,16 +1,14 @@
-package nu.marginalia.util.language.processing;
+package nu.marginalia.util.language.processing.sentence;
 
 import com.github.datquocnguyen.RDRPOSTagger;
 import com.github.jknack.handlebars.internal.lang3.StringUtils;
-import gnu.trove.list.array.TIntArrayList;
 import gnu.trove.map.hash.TObjectIntHashMap;
-import lombok.AllArgsConstructor;
-import lombok.Getter;
 import lombok.SneakyThrows;
+import nu.marginalia.util.StringPool;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.util.language.processing.HtmlTagCleaner;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
-import nu.marginalia.util.language.processing.model.tag.WordSeparator;
 import opennlp.tools.sentdetect.SentenceDetectorME;
 import opennlp.tools.sentdetect.SentenceModel;
 import opennlp.tools.stemmer.PorterStemmer;
@@ -24,25 +22,22 @@ import javax.inject.Inject;
 import java.io.FileInputStream;
 import java.io.IOException;
 import java.io.InputStream;
-import java.util.ArrayList;
-import java.util.Arrays;
-import java.util.List;
-import java.util.Optional;
+import java.util.*;
 import java.util.regex.Pattern;
 
-import static nu.marginalia.util.language.WordPatterns.*;
-
 public class SentenceExtractor {
 
     private SentenceDetectorME sentenceDetector;
     private final RDRPOSTagger rdrposTagger;
 
     private final PorterStemmer porterStemmer = new PorterStemmer();
-    private boolean legacyMode = false;
     private static final Logger logger = LoggerFactory.getLogger(SentenceExtractor.class);
 
     private static final HtmlTagCleaner tagCleaner = new HtmlTagCleaner();
 
+    private final ThreadLocal<StringPool> stringPool = ThreadLocal.withInitial(() -> StringPool.create(10_000));
+
+
     @SneakyThrows @Inject
     public SentenceExtractor(LanguageModels models) {
         try (InputStream modelIn = new FileInputStream(models.openNLPSentenceDetectionData.toFile())) {
@@ -66,6 +61,22 @@ public class SentenceExtractor {
         final String text = asText(doc);
         final DocumentSentence[] textSentences = extractSentencesFromString(text);
 
+        String title = getTitle(doc, textSentences);
+
+        TObjectIntHashMap<String> counts = calculateWordCounts(textSentences);
+        var titleSentences = extractSentencesFromString(title.toLowerCase());
+        return new DocumentLanguageData(textSentences, titleSentences, counts);
+    }
+
+    public DocumentLanguageData extractSentences(String text, String title) {
+        final DocumentSentence[] textSentences = extractSentencesFromString(text);
+
+        TObjectIntHashMap<String> counts = calculateWordCounts(textSentences);
+
+        return new DocumentLanguageData(textSentences, extractSentencesFromString(title.toLowerCase()), counts);
+    }
+
+    private String getTitle(Document doc, DocumentSentence[] textSentences) {
         String title = doc.getElementsByTag("title").text() + " . "  +
                 Optional.ofNullable(doc.getElementsByTag("h1").first()).map(Element::text).orElse("");
 
@@ -82,34 +93,7 @@ public class SentenceExtractor {
             }
         }
 
-        TObjectIntHashMap<String> counts = calculateWordCounts(textSentences);
-        var titleSentences = extractSentencesFromString(title.toLowerCase());
-        return new DocumentLanguageData(textSentences, titleSentences, counts);
-    }
-
-    public DocumentLanguageData extractSentences(String text) {
-        final DocumentSentence[] textSentences = extractSentencesFromString(text);
-
-        String title = "";
-        for (DocumentSentence textSentence : textSentences) {
-            if (textSentence.length() > 0) {
-                title = textSentence.originalSentence.toLowerCase();
-                break;
-            }
-        }
-
-        TObjectIntHashMap<String> counts = calculateWordCounts(textSentences);
-
-        return new DocumentLanguageData(textSentences, extractSentencesFromString(title.toLowerCase()), counts);
-    }
-
-
-    public DocumentLanguageData extractSentences(String text, String title) {
-        final DocumentSentence[] textSentences = extractSentencesFromString(text);
-
-        TObjectIntHashMap<String> counts = calculateWordCounts(textSentences);
-
-        return new DocumentLanguageData(textSentences, extractSentencesFromString(title.toLowerCase()), counts);
+        return title;
     }
 
 
@@ -125,79 +109,95 @@ public class SentenceExtractor {
         return counts;
     }
 
-    private static final Pattern splitPattern = Pattern.compile("( -|- |\\|)");
-
-//    private static final Pattern badCharPattern = Pattern.compile("([^_#@.a-zA-Z'+\\-0-9\\u00C0-\\u00D6\\u00D8-\\u00f6\\u00f8-\\u00ff]+)|(\\.(\\s+|$))");
-
-    private boolean isBadChar(char c) {
-        if (c >= 'a' && c <= 'z') return false;
-        if (c >= 'A' && c <= 'Z') return false;
-        if (c >= '0' && c <= '9') return false;
-        if ("_#@.".indexOf(c) >= 0) return false;
-        if (c >= '\u00C0' && c <= '\u00D6') return false;
-        if (c >= '\u00D8' && c <= '\u00F6') return false;
-        if (c >= '\u00F8' && c <= '\u00FF') return false;
-
-        return true;
-    }
-    private String sanitizeString(String s) {
-        char[] newChars = new char[s.length()];
-        int pi = 0;
-
-        for (int i = 0; i < newChars.length; i++) {
-            char c = s.charAt(i);
-            if (!isBadChar(c)) {
-                newChars[pi++] = c;
-            }
-            else {
-                newChars[pi++] = ' ';
-            }
-        }
-
-        s = new String(newChars, 0, pi);
-
-        if (s.startsWith(".")) {
-            s = s.substring(1);
-            if (s.isBlank())
-                return "";
-        }
-        return s;
-
-    }
-
     public DocumentSentence extractSentence(String text) {
-        var wordsAndSeps = splitSegment(text);
+        var wordsAndSeps = SentenceSegmentSplitter.splitSegment(text);
 
         var words = wordsAndSeps.words;
         var seps = wordsAndSeps.separators;
-        var lc = toLc(wordsAndSeps.words);
+        var lc = SentenceExtractorStringUtils.toLowerCaseStripPossessive(wordsAndSeps.words);
 
         return new DocumentSentence(
-            sanitizeString(text), words, seps, lc, rdrposTagger.tagsForEnSentence(words), stemSentence(lc)
+            SentenceExtractorStringUtils.sanitizeString(text), words, seps, lc, rdrposTagger.tagsForEnSentence(words), stemSentence(lc)
         );
     }
 
-    public String normalizeSpaces(String s) {
-        if (s.indexOf('\t') >= 0) {
-            s = s.replace('\t', ' ');
-        }
-        if (s.indexOf('\n') >= 0) {
-            s = s.replace('\n', ' ');
-        }
-        return s;
-    }
-
     public DocumentSentence[] extractSentencesFromString(String text) {
         String[] sentences;
 
-        String textNormalizedSpaces = normalizeSpaces(text);
+        String textNormalizedSpaces = SentenceExtractorStringUtils.normalizeSpaces(text);
         try {
             sentences = sentenceDetector.sentDetect(textNormalizedSpaces);
         }
         catch (Exception ex) {
+            // shitty fallback logic
             sentences = StringUtils.split(textNormalizedSpaces, '.');
         }
 
+        sentences = preCleanSentences(sentences);
+
+        final String[][] tokens = new String[sentences.length][];
+        final int[][] separators = new int[sentences.length][];
+        final String[][] posTags = new String[sentences.length][];
+        final String[][] tokensLc = new String[sentences.length][];
+        final String[][] stemmedWords = new String[sentences.length][];
+
+        for (int i = 0; i < tokens.length; i++) {
+
+            var wordsAndSeps = SentenceSegmentSplitter.splitSegment(sentences[i]);
+            tokens[i] = wordsAndSeps.words;
+            separators[i] = wordsAndSeps.separators;
+            if (tokens[i].length > 250) {
+                tokens[i] = Arrays.copyOf(tokens[i], 250);
+                separators[i] = Arrays.copyOf(separators[i], 250);
+            }
+            for (int j = 0; j < tokens[i].length; j++) {
+                while (tokens[i][j].endsWith(".")) {
+                    tokens[i][j] = StringUtils.removeEnd(tokens[i][j], ".");
+                }
+            }
+        }
+
+        var sPool = stringPool.get();
+
+        for (int i = 0; i < tokens.length; i++) {
+            tokens[i] = sPool.internalize(tokens[i]);
+        }
+
+        for (int i = 0; i < tokens.length; i++) {
+            posTags[i] = rdrposTagger.tagsForEnSentence(tokens[i]);
+            // don't need to internalize this
+        }
+
+        for (int i = 0; i < tokens.length; i++) {
+            tokensLc[i] = SentenceExtractorStringUtils.toLowerCaseStripPossessive(tokens[i]);
+            tokensLc[i] = sPool.internalize(tokensLc[i]);
+        }
+
+        for (int i = 0; i < tokens.length; i++) {
+            stemmedWords[i] = stemSentence(tokensLc[i]);
+            stemmedWords[i] = sPool.internalize(stemmedWords[i]);
+        }
+
+        DocumentSentence[] ret = new DocumentSentence[sentences.length];
+        for (int i = 0; i < ret.length; i++) {
+            String fullString;
+
+            if (i == 0) {
+                fullString = SentenceExtractorStringUtils.sanitizeString(sentences[i]);
+            }
+            else {
+                fullString = "";
+            }
+
+            ret[i] = new DocumentSentence(fullString, tokens[i], separators[i], tokensLc[i], posTags[i], stemmedWords[i]);
+        }
+        return ret;
+    }
+
+    private static final Pattern splitPattern = Pattern.compile("( -|- |\\|)");
+
+    private String[] preCleanSentences(String[] sentences) {
+
         if (sentences.length > 250) {
             sentences = Arrays.copyOf(sentences, 250);
         }
@@ -212,53 +212,13 @@ public class SentenceExtractor {
                 sentenceList.add(s);
             }
         }
-        sentences = sentenceList.toArray(String[]::new);
-
-        final String[][] tokens = new String[sentences.length][];
-        final int[][] separators = new int[sentences.length][];
-        final String[][] posTags = new String[sentences.length][];
-        final String[][] tokensLc = new String[sentences.length][];
-        final String[][] stemmedWords = new String[sentences.length][];
-
-        for (int i = 0; i < tokens.length; i++) {
-
-            var wordsAndSeps = splitSegment(sentences[i]); //tokenizer.tokenize(sentences[i]);
-            tokens[i] = wordsAndSeps.words;
-            separators[i] = wordsAndSeps.separators;
-            if (tokens[i].length > 250) {
-                tokens[i] = Arrays.copyOf(tokens[i], 250);
-                separators[i] = Arrays.copyOf(separators[i], 250);
-            }
-            for (int j = 0; j < tokens[i].length; j++) {
-                while (tokens[i][j].endsWith(".")) {
-                    tokens[i][j] = StringUtils.removeEnd(tokens[i][j], ".");
-                }
-            }
-        }
-
-        for (int i = 0; i < tokens.length; i++) {
-            posTags[i] = rdrposTagger.tagsForEnSentence(tokens[i]);
-        }
-
-        for (int i = 0; i < tokens.length; i++) {
-            tokensLc[i] = toLc(tokens[i]);
-        }
-
-        for (int i = 0; i < tokens.length; i++) {
-            stemmedWords[i] = stemSentence(tokensLc[i]);
-        }
-
-        DocumentSentence[] ret = new DocumentSentence[sentences.length];
-        for (int i = 0; i < ret.length; i++) {
-            ret[i] = new DocumentSentence(sanitizeString(sentences[i]), tokens[i], separators[i], tokensLc[i], posTags[i], stemmedWords[i]);
-        }
-        return ret;
+        return sentenceList.toArray(String[]::new);
     }
 
     private String[] stemSentence(String[] strings) {
         String[] stemmed = new String[strings.length];
         for (int i = 0; i < stemmed.length; i++) {
-            var sent = cleanPossessive(strings[i]);
+            var sent = SentenceExtractorStringUtils.stripPossessive(strings[i]);
             try {
                 stemmed[i] = porterStemmer.stem(sent);
             }
@@ -269,27 +229,6 @@ public class SentenceExtractor {
         return stemmed;
     }
 
-    private String cleanPossessive(String s) {
-        int end = s.length();
-
-        if (s.endsWith("\'")) {
-            return s.substring(0, end-1);
-        } else if (end > 2 && s.charAt(end-2) == '\'' && "sS".indexOf(s.charAt(end-1))>=0) {
-            return s.substring(0, end-2).toLowerCase();
-        }
-        else {
-            return s;
-        }
-    }
-
-    private String[] toLc(String[] words) {
-        String[] lower = new String[words.length];
-        for (int i = 0; i < lower.length; i++) {
-            lower[i] = cleanPossessive(words[i]).toLowerCase();
-        }
-        return lower;
-    }
-
     public String asText(Document dc) {
 
         tagCleaner.clean(dc);
@@ -299,67 +238,6 @@ public class SentenceExtractor {
         return text.substring(0, (int) (text.length()*0.95));
     }
 
-    @AllArgsConstructor @Getter
-    private static class WordsAndSeparators {
-        String[] words;
-        int[] separators;
-    }
 
-    private WordsAndSeparators splitSegment(String segment) {
-        var matcher = wordBreakPattern.matcher(segment);
-
-        List<String> words = new ArrayList<>(segment.length()/6);
-        TIntArrayList separators = new TIntArrayList(segment.length()/6);
-
-        int start = 0;
-        int wordStart = 0;
-        while (wordStart <= segment.length()) {
-            if (!matcher.find(wordStart)) {
-                words.add(segment.substring(wordStart));
-                separators.add(WordSeparator.SPACE);
-                break;
-            }
-
-            if (wordStart != matcher.start()) {
-                words.add(segment.substring(wordStart, matcher.start()));
-                separators.add(segment.substring(matcher.start(), matcher.end()).isBlank() ? WordSeparator.SPACE : WordSeparator.COMMA);
-            }
-            wordStart = matcher.end();
-        }
-
-        String[] parts = words.toArray(String[]::new);
-        int length = 0;
-        for (int i = 0; i < parts.length; i++) {
-            if (parts[i].isBlank() || parts[i].length() >= MAX_WORD_LENGTH || characterNoisePredicate.test(parts[i])) {
-                parts[i] = null;
-            }
-            else {
-                length++;
-            }
-        }
-
-        String[] ret = new String[length];
-        int[] seps = new int[length];
-        for (int i = 0, j=0; i < parts.length; i++) {
-            if (parts[i] != null) {
-                seps[j] = separators.getQuick(i);
-                ret[j++] = parts[i];
-            }
-        }
-
-        for (int i = 0; i < ret.length; i++) {
-            if (ret[i].startsWith("'") && ret[i].length() > 1) { ret[i] = ret[i].substring(1); }
-            if (ret[i].endsWith("'") && ret[i].length() > 1) { ret[i] = ret[i].substring(0, ret[i].length()-1); }
-        }
-        return new WordsAndSeparators(ret, seps);
-    }
-
-
-    public boolean isLegacyMode() {
-        return legacyMode;
-    }
-    public void setLegacyMode(boolean legacyMode) {
-        this.legacyMode = legacyMode;
-    }
 
 }

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/sentence/SentenceExtractorStringUtils.java

@@ -0,0 +1,93 @@
+package nu.marginalia.util.language.processing.sentence;
+
+import java.util.Arrays;
+import java.util.Objects;
+
+public class SentenceExtractorStringUtils {
+
+    public static String sanitizeString(String s) {
+        char[] newChars = new char[s.length()];
+        int pi = 0;
+        boolean changed = false;
+        for (int i = 0; i < newChars.length; i++) {
+            char c = s.charAt(i);
+            if (!isBadChar(c)) {
+                newChars[pi++] = c;
+            }
+            else {
+                changed = true;
+                newChars[pi++] = ' ';
+            }
+        }
+
+        if (changed) {
+            s = new String(newChars, 0, pi);
+        }
+
+        if (s.startsWith(".")) {
+            s = s.substring(1);
+        }
+
+        if (s.isBlank()) {
+            return "";
+        }
+
+        return s;
+
+    }
+
+    private static boolean isBadChar(char c) {
+        if (c >= 'a' && c <= 'z') return false;
+        if (c >= 'A' && c <= 'Z') return false;
+        if (c >= '0' && c <= '9') return false;
+        if ("_#@.".indexOf(c) >= 0) return false;
+        if (c >= '\u00C0' && c <= '\u00D6') return false;
+        if (c >= '\u00D8' && c <= '\u00F6') return false;
+        if (c >= '\u00F8' && c <= '\u00FF') return false;
+
+        return true;
+    }
+
+    public static String normalizeSpaces(String s) {
+        if (s.indexOf('\t') >= 0) {
+            s = s.replace('\t', ' ');
+        }
+        if (s.indexOf('\n') >= 0) {
+            s = s.replace('\n', ' ');
+        }
+        return s;
+    }
+
+
+    public static String toLowerCaseStripPossessive(String word) {
+        String val = stripPossessive(word).toLowerCase();
+
+        if (Objects.equals(val, word)) {
+            return word;
+        }
+
+        return val;
+    }
+
+    public static String[] toLowerCaseStripPossessive(String[] words) {
+        String[] lc = new String[words.length];
+        Arrays.setAll(lc, i ->SentenceExtractorStringUtils.toLowerCaseStripPossessive(words[i]));
+        return lc;
+    }
+
+    public static String stripPossessive(String s) {
+        int end = s.length();
+
+        if (s.endsWith("'")) {
+            return s.substring(0, end-1);
+        }
+
+        if (s.endsWith("'s") || s.endsWith("'S")) {
+            return s.substring(0, end-2);
+        }
+
+        return s;
+    }
+
+
+}

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/sentence/SentenceSegmentSplitter.java

@@ -0,0 +1,72 @@
+package nu.marginalia.util.language.processing.sentence;
+
+import gnu.trove.list.array.TIntArrayList;
+import lombok.AllArgsConstructor;
+import lombok.Getter;
+import nu.marginalia.util.language.processing.model.tag.WordSeparator;
+
+import java.util.ArrayList;
+import java.util.List;
+
+import static nu.marginalia.util.language.WordPatterns.*;
+
+public class SentenceSegmentSplitter {
+
+
+    @AllArgsConstructor
+    @Getter
+    public static class SeparatedSentence {
+        String[] words;
+        int[] separators;
+    }
+
+    public static SeparatedSentence splitSegment(String segment) {
+        var matcher = wordBreakPattern.matcher(segment);
+
+        List<String> words = new ArrayList<>(segment.length()/6);
+        TIntArrayList separators = new TIntArrayList(segment.length()/6);
+
+        int wordStart = 0;
+        while (wordStart <= segment.length()) {
+            if (!matcher.find(wordStart)) {
+                words.add(segment.substring(wordStart));
+                separators.add(WordSeparator.SPACE);
+                break;
+            }
+
+            if (wordStart != matcher.start()) {
+                words.add(segment.substring(wordStart, matcher.start()));
+                separators.add(segment.substring(matcher.start(), matcher.end()).isBlank() ? WordSeparator.SPACE : WordSeparator.COMMA);
+            }
+            wordStart = matcher.end();
+        }
+
+        String[] parts = words.toArray(String[]::new);
+        int length = 0;
+        for (int i = 0; i < parts.length; i++) {
+            if (parts[i].isBlank() || parts[i].length() >= MAX_WORD_LENGTH || characterNoisePredicate.test(parts[i])) {
+                parts[i] = null;
+            }
+            else {
+                length++;
+            }
+        }
+
+        String[] ret = new String[length];
+        int[] seps = new int[length];
+        for (int i = 0, j=0; i < parts.length; i++) {
+            if (parts[i] != null) {
+                seps[j] = separators.getQuick(i);
+                ret[j++] = parts[i];
+            }
+        }
+
+        for (int i = 0; i < ret.length; i++) {
+            if (ret[i].startsWith("'") && ret[i].length() > 1) { ret[i] = ret[i].substring(1); }
+            if (ret[i].endsWith("'") && ret[i].length() > 1) { ret[i] = ret[i].substring(0, ret[i].length()-1); }
+        }
+        return new SeparatedSentence(ret, seps);
+    }
+
+
+}

marginalia_nu/src/main/java/nu/marginalia/util/ranking/BuggyReversePageRank.java

@@ -1,39 +0,0 @@
-package nu.marginalia.util.ranking;
-
-
-public class BuggyReversePageRank extends RankingAlgorithm {
-
-
-    public BuggyReversePageRank(RankingDomainFetcher domains, String... origins) {
-        super(domains, origins);
-    }
-
-    @Override
-    RankVector createNewRankVector(RankVector rank) {
-
-        double rankNorm = rank.norm();
-        RankVector newRank = new RankVector(0);
-
-        for (int domainId = 0; domainId < domainIndexToId.size(); domainId++) {
-
-            var links = linkDataSrc2Dest[domainId];
-
-            if (links != null && links.size() > 0) {
-                double newRankValue = 0;
-
-                for (int j = 0; j < links.size(); j++) {
-                    newRankValue += rank.get(links.getQuick(j)) / links.size();
-                }
-
-                newRank.set(domainId, 0.85*newRankValue/rankNorm);
-            }
-        }
-        return newRank;
-    }
-
-    @Override
-    void adjustRankVector(RankVector vector, double dNorm, double oldNorm) {
-        originDomainIds.forEach(id -> vector.increment(domainIdToIndex.get(id), dNorm/oldNorm));
-    }
-
-}

marginalia_nu/src/main/java/nu/marginalia/util/ranking/BuggyStandardPageRank.java

@@ -1,45 +0,0 @@
-package nu.marginalia.util.ranking;
-
-
-public class BuggyStandardPageRank extends RankingAlgorithm {
-
-    public BuggyStandardPageRank(RankingDomainFetcher domains, String... origins) {
-        super(domains, origins);
-    }
-
-    @Override
-    RankingAlgorithm.RankVector createNewRankVector(RankingAlgorithm.RankVector rank) {
-        RankVector newRank = new RankVector(0);
-
-        for (int domainId = 0; domainId < domainIndexToId.size(); domainId++) {
-
-            var links = linkDataSrc2Dest[domainId];
-            double newRankValue = 0;
-
-            if (links != null && links.size() > 0) {
-                for (int j = 0; j < links.size(); j++) {
-                    int linkedDomain = links.getQuick(j);
-
-                    int linkSize = 1;
-                    var bl = linkDataSrc2Dest[linkedDomain];
-                    if (bl != null) {
-                        linkSize = bl.size();
-                    }
-
-                    newRankValue += rank.get(linkedDomain) / linkSize;
-
-                }
-            }
-
-            newRank.set(domainId,  0.85 * newRankValue);
-        }
-        return newRank;
-    }
-
-    @Override
-    void adjustRankVector(RankingAlgorithm.RankVector vector, double dNorm, double oldNorm) {
-        originDomainIds.forEach(id -> vector.increment(id, dNorm/originDomainIds.size()));
-        vector.incrementAll(0.14*dNorm/vector.size());
-    }
-
-}

marginalia_nu/src/main/java/nu/marginalia/util/ranking/tool/DedupTool.java

@@ -1,89 +0,0 @@
-package nu.marginalia.util.ranking.tool;
-
-import lombok.AllArgsConstructor;
-import lombok.Getter;
-import lombok.SneakyThrows;
-import lombok.ToString;
-import nu.marginalia.wmsa.configuration.module.DatabaseModule;
-import org.mariadb.jdbc.Driver;
-import org.slf4j.Logger;
-import org.slf4j.LoggerFactory;
-
-import java.io.IOException;
-import java.util.*;
-import java.util.concurrent.LinkedBlockingQueue;
-import java.util.stream.Collectors;
-import java.util.stream.Stream;
-
-public class DedupTool {
-
-    private static final Logger logger = LoggerFactory.getLogger(DedupTool.class);
-
-    public Set<String> originDomains = new HashSet<>();
-    public Set<Integer> originDomainIds = new HashSet<>();
-    public final long domainIdMax = -1;
-    public int domainCount;
-    private volatile static int rankMax;
-
-    public int maxId() {
-        return (int) domainIdMax;
-    }
-    public int domainCount() {
-        return domainCount;
-    }
-
-    static LinkedBlockingQueue<Integer> uploadQueue = new LinkedBlockingQueue<>(10);
-    volatile static boolean running = true;
-
-    @AllArgsConstructor @ToString @Getter
-    static class Data {
-        String url;
-        int id;
-        String domain;
-    }
-
-    @SneakyThrows
-    public static void main(String... args) {
-        Driver driver = new Driver();
-        var ds = new DatabaseModule().provideConnection();
-
-        Map<Integer, Map<Integer, List<Data>>> domainToHashToUrl = new HashMap<>();
-
-        try (var conn = ds.getConnection();
-             var fetchStmt = conn.prepareStatement("SELECT URL_TOP_DOMAIN_ID,DATA_HASH,URL,EC_URL.ID,EC_DOMAIN.DOMAIN_NAME FROM EC_URL INNER JOIN EC_DOMAIN ON EC_DOMAIN.ID=DOMAIN_ID WHERE DATA_HASH IS NOT NULL");
-             var updateStmt = conn.prepareStatement("UPDATE EC_URL SET STATE='redirect' WHERE ID=?")
-
-        ) {
-            fetchStmt.setFetchSize(10_000);
-            var rsp = fetchStmt.executeQuery();
-            while (rsp.next()) {
-                domainToHashToUrl.computeIfAbsent(rsp.getInt(1), i -> new HashMap<>())
-                        .computeIfAbsent(rsp.getInt(2), i -> new ArrayList<>()).add(new Data(rsp.getString(3), rsp.getInt(4), rsp.getString(5)));
-            }
-
-
-            List<Integer> updateIds = new ArrayList<>();
-
-            domainToHashToUrl.forEach((domain, hashes) -> {
-                hashes.forEach((hash, urls) -> {
-                    if (urls.size() > 1) {
-                        Comparator<Data> c = Comparator.comparing(d -> d.domain.length());
-                        var urls2 = urls.stream().sorted(c.thenComparing(d -> d.url.length()))
-                                .collect(Collectors.partitioningBy(d -> d.url.endsWith("/")));
-
-                        Stream
-                                .concat(urls2.get(true).stream(),urls2.get(false).stream()).skip(1)
-                                .map(Data::getId)
-                                .forEach(updateIds::add);
-                    }
-                });
-            });
-
-            for (int id : updateIds) {
-                updateStmt.setInt(1, id);
-                updateStmt.executeUpdate();
-            }
-        }
-    }
-
-}

marginalia_nu/src/main/java/nu/marginalia/wmsa/client/GsonFactory.java

@@ -2,6 +2,7 @@ package nu.marginalia.wmsa.client;
 
 import com.google.gson.*;
 import marcono1234.gson.recordadapter.RecordTypeAdapterFactory;
+import nu.marginalia.util.bigstring.BigString;
 import nu.marginalia.wmsa.edge.model.EdgeDomain;
 import nu.marginalia.wmsa.edge.model.EdgeUrl;
 import nu.marginalia.wmsa.edge.model.id.EdgeId;
@@ -24,6 +25,8 @@ public class GsonFactory {
                 .registerTypeAdapter(EdgeDomain.class, (JsonDeserializer<EdgeDomain>) (json, typeOfT, context) -> new EdgeDomain(json.getAsString()))
                 .registerTypeAdapter(EdgeId.class, (JsonDeserializer<EdgeId<?>>) (json, typeOfT, context) -> new EdgeId<>(json.getAsInt()))
                 .registerTypeAdapter(EdgeId.class, (JsonSerializer<EdgeId<?>>) (src, typeOfSrc, context) -> new JsonPrimitive(src.id()))
+                .registerTypeAdapter(BigString.class, (JsonDeserializer<BigString>) (json, typeOfT, context) -> BigString.encode(json.getAsString()))
+                .registerTypeAdapter(BigString.class, (JsonSerializer<BigString>) (src, typeOfT, context) -> new JsonPrimitive(src.decode()))
                 .serializeSpecialFloatingPointValues()
                 .create();
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/configuration/ServiceDescriptor.java

@@ -13,7 +13,6 @@ import nu.marginalia.wmsa.memex.MemexMain;
 import nu.marginalia.wmsa.podcasts.PodcastScraperMain;
 import nu.marginalia.wmsa.renderer.RendererMain;
 import nu.marginalia.wmsa.resource_store.ResourceStoreMain;
-import nu.marginalia.wmsa.smhi.scraper.SmhiScraperMain;
 import org.apache.logging.log4j.core.lookup.MainMapLookup;
 
 import java.util.Map;
@@ -26,7 +25,6 @@ public enum ServiceDescriptor {
     AUTH("auth", 5003, AuthMain.class),
     API("api", 5004, ApiMain.class),
 
-    SMHI_SCRAPER("smhi-scraper",5012, SmhiScraperMain.class),
     PODCST_SCRAPER("podcast-scraper", 5013, PodcastScraperMain.class),
 
     EDGE_INDEX("edge-index", 5021, EdgeIndexMain.class),

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/assistant/dict/TermFrequencyDict.java

@@ -2,9 +2,10 @@ package nu.marginalia.wmsa.edge.assistant.dict;
 
 import ca.rmen.porterstemmer.PorterStemmer;
 import gnu.trove.map.hash.TLongIntHashMap;
+import gnu.trove.set.hash.TLongHashSet;
 import nu.marginalia.util.language.LanguageFilter;
 import nu.marginalia.util.language.conf.LanguageModels;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.wmsa.configuration.WmsaHome;
 import nu.marginalia.wmsa.edge.converting.processor.logic.DomPruningFilter;
@@ -18,11 +19,10 @@ import javax.annotation.Nullable;
 import javax.inject.Inject;
 import javax.inject.Singleton;
 import java.io.*;
+import java.nio.charset.StandardCharsets;
 import java.nio.file.Files;
 import java.nio.file.Path;
 import java.util.Arrays;
-import java.util.HashSet;
-import java.util.Set;
 import java.util.concurrent.ForkJoinPool;
 import java.util.concurrent.TimeUnit;
 import java.util.concurrent.atomic.AtomicInteger;
@@ -101,12 +101,15 @@ public class TermFrequencyDict {
 
             fjp.execute(() -> {
 
+                TLongHashSet words = new TLongHashSet(10_000);
+
                 for (var doc : domain.doc) {
+
                     if (doc.documentBody == null)
                         continue;
                     docCount.incrementAndGet();
 
-                    Document parsed = Jsoup.parse(doc.documentBody);
+                    Document parsed = Jsoup.parse(doc.documentBody.decode());
                     parsed.body().filter(new DomPruningFilter(0.5));
 
                     DocumentLanguageData dld = se.get().extractSentences(parsed);
@@ -115,28 +118,30 @@ public class TermFrequencyDict {
                         return;
                     }
 
-                    Set<String> words = new HashSet<>(10_000);
-
                     for (var sent : dld.sentences) {
                         for (var word : sent) {
-                            words.add(word.stemmed());
+                            words.add(longHash(word.stemmed().getBytes(StandardCharsets.UTF_8)));
                         }
                     }
 
-                    fjp.execute(() -> {
-                        synchronized (counts) {
-                            for (var word : words) {
-                                counts.adjustOrPutValue(longHash(word.getBytes()),  1, 1);
-                            }
-                        }
-                    });
+                    synchronized (counts) {
+                        words.forEach(w -> {
+                            counts.adjustOrPutValue(w,  1, 1);
+                            return true;
+                        });
+                    }
 
+                    words.clear();
                 }
+
+                System.out.println(domain.domain + "\t" + counts.size());
             });
+
+
         }
 
         fjp.shutdown();
-        fjp.awaitTermination(10, TimeUnit.SECONDS);
+        fjp.awaitTermination(10, TimeUnit.DAYS);
 
         try (var dos = new DataOutputStream(Files.newOutputStream(Path.of(outFile)))) {
             synchronized (counts) {
@@ -155,14 +160,6 @@ public class TermFrequencyDict {
         }
 
         System.out.println(docCount.get());
-//
-//        counts.forEachEntry((w,c) -> {
-//            if (c > 3L) {
-//                System.out.println(w + ":" + c);
-//            }
-//            return true;
-//        });
-
     }
 
     public static long getStringHash(String s) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/ConverterMain.java

@@ -46,17 +46,12 @@ public class ConverterMain {
             InstructionsCompiler compiler,
             Gson gson
             ) throws Exception {
-
-        ;
-
-
-
         logger.info("Starting pipe");
 
         try (WorkLog processLog = plan.createProcessWorkLog();
              ConversionLog log = new ConversionLog(plan.process.getDir())) {
             instructionWriter = new LoadInstructionWriter(log, plan.process.getDir(), gson);
-            var pipe = new ParallelPipe<CrawledDomain, ProcessingInstructions>("Crawler", 20, 4, 2) {
+            var pipe = new ParallelPipe<CrawledDomain, ProcessingInstructions>("Crawler", 16, 4, 2) {
 
                 @Override
                 protected ProcessingInstructions onProcess(CrawledDomain domainData) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/LinkKeywordExtractorMain.java

@@ -150,7 +150,7 @@ public class LinkKeywordExtractorMain {
 
                 for (var doc : crawledDomain.doc) {
                     if (Objects.equals(doc.crawlerStatus, CrawlerDocumentStatus.OK.name())) {
-                        anchorTextExtractor.processDocument(doc.url, doc.documentBody);
+                        anchorTextExtractor.processDocument(doc.url, doc.documentBody.decode());
                     }
                 }
             }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/DocumentProcessor.java

@@ -7,7 +7,7 @@ import nu.marginalia.util.gregex.GuardedRegex;
 import nu.marginalia.util.gregex.GuardedRegexFactory;
 import nu.marginalia.util.language.LanguageFilter;
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.util.language.processing.model.KeywordMetadata;
 import nu.marginalia.wmsa.edge.converting.model.DisqualifiedException;
@@ -178,11 +178,13 @@ public class DocumentProcessor {
     private DetailsWithWords createDetails(CrawledDomain crawledDomain, CrawledDocument crawledDocument)
             throws DisqualifiedException, URISyntaxException {
 
-        if (languageFilter.isBlockedUnicodeRange(crawledDocument.documentBody)) {
+        String documentBody = crawledDocument.documentBody.decode();
+
+        if (languageFilter.isBlockedUnicodeRange(documentBody)) {
             throw new DisqualifiedException(DisqualificationReason.LANGUAGE);
         }
 
-        Document doc = Jsoup.parse(crawledDocument.documentBody);
+        Document doc = Jsoup.parse(documentBody);
 
         if (AcceptableAds.hasAcceptableAdsTag(doc)) {
             // I've never encountered a website where this hasn't been a severe indicator

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/DomainProcessor.java

@@ -42,7 +42,7 @@ public class DomainProcessor {
 
             fixBadCanonicalTags(crawledDomain.doc);
 
-            StringPool stringPool = new StringPool(1000 + 100 * crawledDomain.doc.size());
+            StringPool stringPool = StringPool.create(1000 + 100 * crawledDomain.doc.size());
 
             for (var doc : crawledDomain.doc) {
                 var processedDoc = documentProcessor.process(doc, crawledDomain);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/SiteWords.java

@@ -33,8 +33,7 @@ public class SiteWords {
         Set<String> commonSiteWords = new HashSet<>(10);
 
         commonSiteWords.addAll(commonKeywordExtractor.getCommonSiteWords(processedDomain,
-                EdgePageWordFlags.Subjects,
-                EdgePageWordFlags.TfIdfHigh));
+                EdgePageWordFlags.Subjects));
 
         commonSiteWords.addAll(commonKeywordExtractor.getCommonSiteWords(processedDomain,
                 EdgePageWordFlags.Title));

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/CommonKeywordExtractor.java

@@ -11,7 +11,7 @@ public class CommonKeywordExtractor {
 
     private static final int MIN_REQUIRED_DOCUMENTS = 25;
 
-    private static final int REQUIRED_TOTAL_COUNT_FOR_CONSIDERATION = 100;
+    private static final int REQUIRED_TOTAL_COUNT_FOR_CONSIDERATION = 15;
     private static final double QUALIFYING_PROPORTION_FOR_KEYWORD = .25;
 
     private static final int MAX_SITE_KEYWORDS_TO_EXTRACT = 5;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/LinkParser.java

@@ -126,6 +126,9 @@ public class LinkParser {
         if (doesUrlStringHaveProtocol(s)) {
             return s;
         }
+        else if (s.startsWith("//")) { // scheme-relative URL
+            return baseUrl.proto + ":" + s;
+        }
 
         String[] parts = paramSeparatorPattern.split(s, 2);
         String path = parts[0];

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/crawling/CrawledDomainReader.java

@@ -2,11 +2,11 @@ package nu.marginalia.wmsa.edge.crawling;
 
 import com.github.luben.zstd.ZstdInputStream;
 import com.google.gson.Gson;
+import jdkoverride.LargeLineBufferedReader;
 import nu.marginalia.wmsa.client.GsonFactory;
 import nu.marginalia.wmsa.edge.crawling.model.CrawledDocument;
 import nu.marginalia.wmsa.edge.crawling.model.CrawledDomain;
 
-import java.io.BufferedReader;
 import java.io.FileInputStream;
 import java.io.IOException;
 import java.io.InputStreamReader;
@@ -19,61 +19,41 @@ import java.util.concurrent.TimeUnit;
 public class CrawledDomainReader {
     private final Gson gson = GsonFactory.get();
 
-    private final ForkJoinPool pool = new ForkJoinPool(4);
+    private final ForkJoinPool pool = new ForkJoinPool(6);
 
     public CrawledDomainReader() {
     }
 
     public CrawledDomain read(Path path) throws IOException {
-        List<CrawledDocument> docs = new ArrayList<>();
-        CrawledDomain domain = null;
+        DomainDataAssembler domainData = new DomainDataAssembler();
 
+        try (var br = new LargeLineBufferedReader(new InputStreamReader(new ZstdInputStream(new FileInputStream(path.toFile()))))) {
+            String line;
+            while ((line = br.readLine()) != null) {
+                if (line.startsWith("//")) {
+                    String identifier = line;
+                    String data = br.readLine();
 
-        try (var br = new BufferedReader(new InputStreamReader(new ZstdInputStream(new FileInputStream(path.toFile()))))) {
-            br.mark(2);
-            boolean legacy = '{' == br.read();
-            br.reset();
-
-            if (legacy) {
-                domain = gson.fromJson(br, CrawledDomain.class);
-            }
-            else {
-                String line;
-                while ((line = br.readLine()) != null) {
-                    if (line.startsWith("//")) {
-                        String nextLine = br.readLine();
-                        if (nextLine == null) break;
-
-                        if (line.equals(CrawledDomain.SERIAL_IDENTIFIER)) {
-                            domain = gson.fromJson(nextLine, CrawledDomain.class);
-                        } else if (line.equals(CrawledDocument.SERIAL_IDENTIFIER)) {
-                            pool.execute(() -> {
-                                var doc = gson.fromJson(nextLine, CrawledDocument.class);
-                                synchronized (docs) {
-                                    docs.add(doc);
-                                }
-                            });
-                        }
-                    } else if (line.charAt(0) == '{') {
-                        domain = gson.fromJson(line, CrawledDomain.class);
-                    }
+                    pool.execute(() -> deserializeLine(identifier, data, domainData));
                 }
             }
         }
 
-        pool.awaitQuiescence(10, TimeUnit.SECONDS);
+        while (!pool.awaitQuiescence(1, TimeUnit.SECONDS));
 
-        if (domain == null) {
-            return null;
+        return domainData.assemble();
+    }
+
+
+    private void deserializeLine(String identifier, String data, DomainDataAssembler assembler) {
+        if (null == data) {
+            return;
         }
-
-        if (!docs.isEmpty()) {
-            if (domain.doc == null)
-                domain.doc = new ArrayList<>();
-
-            domain.doc.addAll(docs);
+        if (identifier.equals(CrawledDomain.SERIAL_IDENTIFIER)) {
+            assembler.acceptDomain(gson.fromJson(data, CrawledDomain.class));
+        } else if (identifier.equals(CrawledDocument.SERIAL_IDENTIFIER)) {
+            assembler.acceptDoc(gson.fromJson(data, CrawledDocument.class));
         }
-        return domain;
     }
 
     public CrawledDomain readRuntimeExcept(Path path) {
@@ -84,4 +64,27 @@ public class CrawledDomainReader {
             throw new RuntimeException(ex);
         }
     }
+
+    private static class DomainDataAssembler {
+        private CrawledDomain domainPrototype;
+        private final List<CrawledDocument> docs = new ArrayList<>();
+
+        public synchronized void acceptDomain(CrawledDomain domain) {
+            this.domainPrototype = domain;
+        }
+
+        public synchronized void acceptDoc(CrawledDocument doc) {
+            docs.add(doc);
+        }
+
+        public synchronized CrawledDomain assemble() {
+            if (!docs.isEmpty()) {
+                if (domainPrototype.doc == null)
+                    domainPrototype.doc = new ArrayList<>();
+
+                domainPrototype.doc.addAll(docs);
+            }
+            return domainPrototype;
+        }
+    }
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/crawling/blocklist/UrlBlocklist.java

@@ -24,7 +24,7 @@ public class UrlBlocklist {
         patterns.add(s -> s.contains("-download-free"));
 
         // long base64-strings in URLs are typically git hashes or the like, rarely worth crawling
-        patterns.add(GuardedRegexFactory.minLength(48, ".*/[^/]*[a-f0-9]{32,}(/|$)"));
+        patterns.add(this::hashTest);
 
         // link farms &c
         patterns.add(GuardedRegexFactory.contains("/download", "/download(-([A-Za-z]+|[0-9]+)){4,}\\.(htm|html|php)$"));
@@ -38,6 +38,33 @@ public class UrlBlocklist {
 
     }
 
+    public boolean hashTest(String path) {
+        // look for strings might be a git hash (i.e. long hexadecimal strings)
+        // there is no good guard for a regular expression for this so hand-rolling this
+        // is necessary
+
+        int runLength = 0;
+        int minLength = 32;
+
+        if (path.length() <= minLength + 2)
+            return false;
+
+        for (int i = 0; i < path.length(); i++) {
+            int c = path.charAt(i);
+
+            if ((c >= '0' && c <= '9') || (c >= 'a' && c <= 'f')) {
+                runLength++;
+            }
+            else if (runLength >= minLength) {
+                return true;
+            }
+            else {
+                runLength = 0;
+            }
+        }
+        return runLength >= minLength;
+    }
+
     public boolean isUrlBlocked(EdgeUrl url) {
         try {
             if (badDomains.contains(url.domain.domain)) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/crawling/model/CrawledDocument.java

@@ -1,6 +1,8 @@
 package nu.marginalia.wmsa.edge.crawling.model;
 
 import lombok.Builder;
+import nu.marginalia.util.bigstring.BigString;
+import nu.marginalia.util.bigstring.CompressedBigString;
 
 @Builder
 public class CrawledDocument implements SerializableCrawlData {
@@ -16,8 +18,7 @@ public class CrawledDocument implements SerializableCrawlData {
     public String crawlerStatusDesc;
 
     public String headers;
-    public String documentBody;
-
+    public BigString documentBody;
     public String documentBodyHash;
 
     public String canonicalUrl;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/crawling/retreival/CrawlerRetreiver.java

@@ -206,7 +206,7 @@ public class CrawlerRetreiver {
 
             if (doc.documentBody != null) {
 
-                doc.documentBodyHash = createHash(doc.documentBody);
+                doc.documentBodyHash = createHash(doc.documentBody.decode());
 
                 Optional<Document> parsedDoc = parseDoc(doc);
                 EdgeUrl url = new EdgeUrl(doc.url);
@@ -251,7 +251,7 @@ public class CrawlerRetreiver {
     private Optional<Document> parseDoc(CrawledDocument doc) {
         if (doc.documentBody == null)
             return Optional.empty();
-        return Optional.of(Jsoup.parse(doc.documentBody));
+        return Optional.of(Jsoup.parse(doc.documentBody.decode()));
     }
 
     public boolean isSameDomain(EdgeUrl url) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/crawling/retreival/HttpFetcher.java

@@ -7,6 +7,7 @@ import crawlercommons.robots.SimpleRobotRulesParser;
 import lombok.AllArgsConstructor;
 import lombok.SneakyThrows;
 import lombok.ToString;
+import nu.marginalia.util.bigstring.BigString;
 import nu.marginalia.wmsa.edge.crawling.model.CrawledDocument;
 import nu.marginalia.wmsa.edge.crawling.model.CrawlerDocumentStatus;
 import nu.marginalia.wmsa.edge.crawling.retreival.logic.ContentTypeLogic;
@@ -271,7 +272,7 @@ public class HttpFetcher {
                 .canonicalUrl(canonical)
                 .httpStatus(rsp.code())
                 .url(responseUrl.toString())
-                .documentBody(strData)
+                .documentBody(BigString.encode(strData))
                 .build();
     }
 
@@ -325,7 +326,7 @@ public class HttpFetcher {
 
     private SimpleRobotRules parseRobotsTxt(CrawledDocument doc) {
         return robotsParser.parseContent(doc.url,
-                doc.documentBody.getBytes(StandardCharsets.UTF_8),
+                doc.documentBody.getBytes(),
                 doc.contentType,
                 userAgent);
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/EdgeIndexControl.java

@@ -2,6 +2,7 @@ package nu.marginalia.wmsa.edge.index;
 
 
 import com.google.inject.Inject;
+import nu.marginalia.wmsa.edge.index.svc.EdgeIndexSearchSetsService;
 
 import java.io.IOException;
 
@@ -9,14 +10,16 @@ import java.io.IOException;
 public class EdgeIndexControl {
 
     private final IndexServicesFactory servicesFactory;
+    private final EdgeIndexSearchSetsService searchSetsService;
 
     @Inject
-    public EdgeIndexControl(IndexServicesFactory servicesFactory) {
+    public EdgeIndexControl(IndexServicesFactory servicesFactory, EdgeIndexSearchSetsService searchSetsService) {
         this.servicesFactory = servicesFactory;
+        this.searchSetsService = searchSetsService;
     }
 
     public void regenerateIndex() throws IOException {
-        servicesFactory.convertIndex();
+        servicesFactory.convertIndex(searchSetsService.getDomainRankings());
 
         System.gc();
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/EdgeIndexModule.java

@@ -13,13 +13,6 @@ public class EdgeIndexModule extends AbstractModule {
 
 
     public void configure() {
-        if (Boolean.getBoolean("small-ram")) {
-            bind(Long.class).annotatedWith(Names.named("edge-dictionary-hash-map-size")).toInstance(1L << 27);
-        }
-        else {
-            bind(Long.class).annotatedWith(Names.named("edge-dictionary-hash-map-size")).toInstance(1L << 31);
-        }
-
     }
 
     @Provides

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/IndexServicesFactory.java

@@ -5,11 +5,11 @@ import com.google.inject.Singleton;
 import com.google.inject.name.Named;
 import lombok.SneakyThrows;
 import nu.marginalia.util.array.LongArray;
-import nu.marginalia.util.dict.DictionaryHashMap;
-import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklist;
+import nu.marginalia.util.dict.DictionaryMap;
 import nu.marginalia.wmsa.edge.index.lexicon.KeywordLexicon;
 import nu.marginalia.wmsa.edge.index.lexicon.KeywordLexiconReadOnlyView;
 import nu.marginalia.wmsa.edge.index.lexicon.journal.KeywordLexiconJournal;
+import nu.marginalia.wmsa.edge.index.postings.DomainRankings;
 import nu.marginalia.wmsa.edge.index.postings.SearchIndex;
 import nu.marginalia.wmsa.edge.index.postings.SearchIndexReader;
 import nu.marginalia.wmsa.edge.index.postings.forward.ForwardIndexConverter;
@@ -20,6 +20,7 @@ import nu.marginalia.wmsa.edge.index.postings.reverse.ReverseIndexConverter;
 import nu.marginalia.wmsa.edge.index.postings.reverse.ReverseIndexPrioReader;
 import nu.marginalia.wmsa.edge.index.postings.reverse.ReverseIndexPriorityParameters;
 import nu.marginalia.wmsa.edge.index.postings.reverse.ReverseIndexReader;
+import nu.marginalia.wmsa.edge.index.svc.EdgeIndexSearchSetsService;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -33,7 +34,6 @@ import java.util.concurrent.Callable;
 @Singleton
 public class IndexServicesFactory {
     private final Path tmpFileDir;
-    private final EdgeDomainBlacklist domainBlacklist;
 
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
@@ -48,7 +48,6 @@ public class IndexServicesFactory {
     private final PartitionedDataFile revPrioIndexWords;
 
     private volatile static KeywordLexicon keywordLexicon;
-    private final Long dictionaryHashMapSize;
 
     private final Path searchSetsBase;
 
@@ -59,14 +58,10 @@ public class IndexServicesFactory {
     public IndexServicesFactory(
             @Named("tmp-file-dir") Path tmpFileDir,
             @Named("partition-root-slow") Path partitionRootSlow,
-            @Named("partition-root-fast") Path partitionRootFast,
-            @Named("edge-dictionary-hash-map-size") Long dictionaryHashMapSize,
-            EdgeDomainBlacklist domainBlacklist
+            @Named("partition-root-fast") Path partitionRootFast
             ) throws IOException {
 
         this.tmpFileDir = tmpFileDir;
-        this.dictionaryHashMapSize = dictionaryHashMapSize;
-        this.domainBlacklist = domainBlacklist;
 
         this.writerIndexFile = new PartitionedDataFile(partitionRootSlow, "page-index.dat");
         this.keywordLexiconFile = new RootDataFile(partitionRootSlow, "dictionary.dat");
@@ -98,7 +93,7 @@ public class IndexServicesFactory {
     public KeywordLexicon getKeywordLexicon() {
         if (keywordLexicon == null) {
             final var journal = new KeywordLexiconJournal(keywordLexiconFile.get());
-            keywordLexicon = new KeywordLexicon(journal, new DictionaryHashMap(dictionaryHashMapSize));
+            keywordLexicon = new KeywordLexicon(journal, DictionaryMap.create());
         }
         return keywordLexicon;
     }
@@ -109,15 +104,15 @@ public class IndexServicesFactory {
 
     }
 
-    public void convertIndex() throws IOException {
-        convertForwardIndex();
-        convertFullReverseIndex();
-        convertPriorityReverseIndex();
+    public void convertIndex(DomainRankings domainRankings) throws IOException {
+        convertForwardIndex(domainRankings);
+        convertFullReverseIndex(domainRankings);
+        convertPriorityReverseIndex(domainRankings);
 
 
     }
 
-    private void convertFullReverseIndex() throws IOException {
+    private void convertFullReverseIndex(DomainRankings domainRankings) throws IOException {
 
         logger.info("Converting full reverse index");
 
@@ -125,6 +120,7 @@ public class IndexServicesFactory {
         var journalReader = new SearchIndexJournalReaderSingleFile(longArray);
         var converter = new ReverseIndexConverter(tmpFileDir,
                 journalReader,
+                domainRankings,
                 revIndexWords.get(NEXT_PART).toPath(),
                 revIndexDoc.get(NEXT_PART).toPath());
 
@@ -133,7 +129,7 @@ public class IndexServicesFactory {
         tryGc();
     }
 
-    private void convertPriorityReverseIndex() throws IOException {
+    private void convertPriorityReverseIndex(DomainRankings domainRankings) throws IOException {
 
         logger.info("Converting priority reverse index");
 
@@ -143,6 +139,7 @@ public class IndexServicesFactory {
 
         var converter = new ReverseIndexConverter(tmpFileDir,
                 journalReader,
+                domainRankings,
                 revPrioIndexWords.get(NEXT_PART).toPath(),
                 revPrioIndexDoc.get(NEXT_PART).toPath());
 
@@ -151,13 +148,14 @@ public class IndexServicesFactory {
         tryGc();
     }
 
-    private void convertForwardIndex() throws IOException {
+    private void convertForwardIndex(DomainRankings domainRankings) throws IOException {
         logger.info("Converting forward index data");
 
-        new ForwardIndexConverter(tmpFileDir,
+        new ForwardIndexConverter(
                 writerIndexFile.get(0),
                 fwdIndexDocId.get(NEXT_PART).toPath(),
-                fwdIndexDocData.get(NEXT_PART).toPath())
+                fwdIndexDocData.get(NEXT_PART).toPath(),
+                domainRankings)
                 .convert();
 
         tryGc();
@@ -215,8 +213,8 @@ public class IndexServicesFactory {
         }
     }
 
-    public SearchIndex createIndexBucket() {
-        return new SearchIndex(this, new EdgeIndexControl(this));
+    public SearchIndex createIndexBucket(EdgeIndexSearchSetsService searchSetsService) {
+        return new SearchIndex(this, new EdgeIndexControl(this, searchSetsService));
     }
 
     public SearchIndexReader getSearchIndexReader() throws IOException {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/client/EdgeIndexLocalService.java

@@ -3,7 +3,8 @@ package nu.marginalia.wmsa.edge.index.client;
 import com.google.inject.Inject;
 import com.google.inject.Singleton;
 import com.google.inject.name.Named;
-import nu.marginalia.util.dict.DictionaryHashMap;
+import nu.marginalia.util.dict.OffHeapDictionaryHashMap;
+import nu.marginalia.util.dict.DictionaryMap;
 import nu.marginalia.wmsa.configuration.server.Context;
 import nu.marginalia.wmsa.edge.converting.interpreter.instruction.DocumentKeywords;
 import nu.marginalia.wmsa.edge.converting.interpreter.instruction.KeywordListChunker;
@@ -32,14 +33,9 @@ public class EdgeIndexLocalService implements EdgeIndexWriterClient {
 
     @Inject
     public EdgeIndexLocalService(@Named("local-index-path") Path path) throws IOException {
-        long hashMapSize = 1L << 31;
-
-        if (Boolean.getBoolean("small-ram")) {
-            hashMapSize = 1L << 27;
-        }
 
         var lexiconJournal = new KeywordLexiconJournal(path.resolve("dictionary.dat").toFile());
-        lexicon = new KeywordLexicon(lexiconJournal, new DictionaryHashMap(hashMapSize));
+        lexicon = new KeywordLexicon(lexiconJournal, DictionaryMap.create());
         indexWriter = new SearchIndexJournalWriterImpl(lexicon, path.resolve("index.dat").toFile());
     }
 
@@ -72,7 +68,7 @@ public class EdgeIndexLocalService implements EdgeIndexWriterClient {
             String word = words[i];
 
             long id = lexicon.getOrInsert(word);
-            if (id != DictionaryHashMap.NO_VALUE) {
+            if (id != OffHeapDictionaryHashMap.NO_VALUE) {
                 ids[putIdx++] = id;
                 ids[putIdx++] = meta[i];
             }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/lexicon/KeywordLexicon.java

@@ -4,7 +4,6 @@ import com.google.common.hash.HashFunction;
 import com.google.common.hash.Hashing;
 import io.prometheus.client.Gauge;
 import lombok.SneakyThrows;
-import nu.marginalia.util.dict.DictionaryHashMap;
 import nu.marginalia.util.dict.DictionaryMap;
 import nu.marginalia.wmsa.edge.index.lexicon.journal.KeywordLexiconJournal;
 import org.slf4j.Logger;
@@ -55,7 +54,7 @@ public class KeywordLexicon implements AutoCloseable {
     private int getOrInsert(byte[] bytes) {
         if (bytes.length >= Byte.MAX_VALUE) {
             logger.warn("getOrInsert({}), illegal length {}", new String(bytes), bytes.length);
-            return DictionaryHashMap.NO_VALUE;
+            return DictionaryMap.NO_VALUE;
         }
 
         final long key = hashFunction.hashBytes(bytes).padToLong();

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/model/EdgePageDocumentsMetadata.java

@@ -8,7 +8,8 @@ import java.util.Set;
 import static java.lang.Math.max;
 import static java.lang.Math.min;
 
-public record EdgePageDocumentsMetadata(int encSize,
+public record EdgePageDocumentsMetadata(int rank,
+                                        int encSize,
                                         int topology,
                                         int year,
                                         int sets,
@@ -16,9 +17,13 @@ public record EdgePageDocumentsMetadata(int encSize,
                                         byte flags) {
 
 
+    public static final long RANK_MASK = 0xFFL;
+    public static final int RANK_SHIFT = 48;
+
     public static final long ENCSIZE_MASK = 0xFFL;
-    public static final int ENCSIZE_SHIFT = 48;
+    public static final int ENCSIZE_SHIFT = 40;
     public static final int ENCSIZE_MULTIPLIER = 50;
+
     public static final long TOPOLOGY_MASK = 0xFFL;
 
     public static final int TOPOLOGY_SHIFT = 32;
@@ -39,7 +44,7 @@ public record EdgePageDocumentsMetadata(int encSize,
         this(defaultValue());
     }
     public EdgePageDocumentsMetadata(int topology, int year, int sets, int quality, EnumSet<EdgePageDocumentFlags> flags) {
-        this(0, topology, year, sets, quality, encodeFlags(flags));
+        this(0, 0, topology, year, sets, quality, encodeFlags(flags));
     }
 
     public EdgePageDocumentsMetadata withSize(int size) {
@@ -49,7 +54,7 @@ public record EdgePageDocumentsMetadata(int encSize,
 
         final int encSize = (int) Math.min(ENCSIZE_MASK, Math.max(1, size / ENCSIZE_MULTIPLIER));
 
-        return new EdgePageDocumentsMetadata(encSize, topology, year, sets, quality, flags);
+        return new EdgePageDocumentsMetadata(rank, encSize, topology, year, sets, quality, flags);
     }
 
     private static byte encodeFlags(Set<EdgePageDocumentFlags> flags) {
@@ -63,7 +68,8 @@ public record EdgePageDocumentsMetadata(int encSize,
     }
 
     public EdgePageDocumentsMetadata(long value) {
-        this(   (int) ((value >>> ENCSIZE_SHIFT) & ENCSIZE_MASK),
+        this(   (int) ((value >>> RANK_SHIFT) & RANK_MASK),
+                (int) ((value >>> ENCSIZE_SHIFT) & ENCSIZE_MASK),
                 (int) ((value >>> TOPOLOGY_SHIFT) & TOPOLOGY_MASK),
                 (int) ((value >>> YEAR_SHIFT) & YEAR_MASK),
                 (int) ((value >>> SETS_SHIFT) & SETS_MASK),
@@ -84,12 +90,13 @@ public record EdgePageDocumentsMetadata(int encSize,
         ret |= min(YEAR_MASK, max(0, year)) << YEAR_SHIFT;
         ret |= min(TOPOLOGY_MASK, max(0, topology)) << TOPOLOGY_SHIFT;
         ret |= min(ENCSIZE_MASK, max(0, encSize)) << ENCSIZE_SHIFT;
+        ret |= min(RANK_MASK, max(0, rank)) << RANK_SHIFT;
 
         return ret;
     }
 
     public boolean isEmpty() {
-        return encSize == 0 && topology == 0 && sets == 0 && quality == 0 && year == 0 && flags == 0;
+        return encSize == 0 && topology == 0 && sets == 0 && quality == 0 && year == 0 && flags == 0 && rank == 0;
     }
 
     public static int decodeQuality(long encoded) {
@@ -112,6 +119,12 @@ public record EdgePageDocumentsMetadata(int encSize,
         return ENCSIZE_MULTIPLIER * (int) ((encoded >>> ENCSIZE_SHIFT) & ENCSIZE_MASK);
     }
 
+    public static int decodeRank(long encoded) {
+        return  (int) ((encoded >>> RANK_SHIFT) & RANK_MASK);
+    }
 
+    public static long encodeRank(long encoded, int rank) {
+        return encoded | min(RANK_MASK, max(0, rank)) << RANK_SHIFT;
+    }
 
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/model/QueryLimits.java

@@ -0,0 +1,4 @@
+package nu.marginalia.wmsa.edge.index.model;
+
+public record QueryLimits(int resultsByDomain, int resultsTotal, int timeoutMs, int fetchSize) {
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/model/QueryStrategy.java

@@ -3,5 +3,10 @@ package nu.marginalia.wmsa.edge.index.model;
 public enum QueryStrategy {
     SENTENCE,
     TOPIC,
+
+    REQUIRE_FIELD_SITE,
+    REQUIRE_FIELD_TITLE,
+    REQUIRE_FIELD_SUBJECT,
+
     AUTO
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/DomainRankings.java

@@ -0,0 +1,43 @@
+package nu.marginalia.wmsa.edge.index.postings;
+
+import it.unimi.dsi.fastutil.ints.Int2IntOpenHashMap;
+import it.unimi.dsi.fastutil.ints.Int2ShortOpenHashMap;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import static java.lang.Math.max;
+import static java.lang.Math.min;
+
+public class DomainRankings {
+    private final Int2ShortOpenHashMap rankings;
+
+    private final int MAX_MEANINGFUL_RANK = 50_000;
+    private final int MAX_RANK_VALUE = 255;
+    private final int MIN_RANK_VALUE = 1;
+    private final double RANK_SCALING_FACTOR = (double) MAX_RANK_VALUE / MAX_MEANINGFUL_RANK;
+
+    public DomainRankings() {
+        rankings = new Int2ShortOpenHashMap();
+    }
+    public DomainRankings(Int2IntOpenHashMap values) {
+         rankings = new Int2ShortOpenHashMap(values.size());
+         values.forEach(this::putRanking);
+    }
+
+    private void putRanking(int domainId, int value) {
+        rankings.put(domainId, scaleRank(value));
+    }
+
+    private short scaleRank(int value) {
+        double rankScaled = RANK_SCALING_FACTOR * value;
+        return (short) min(MAX_RANK_VALUE, max(MIN_RANK_VALUE, rankScaled));
+    }
+
+    public int getRanking(int domainId) {
+        return rankings.getOrDefault(domainId, (short) MAX_RANK_VALUE);
+    }
+
+    public int size() {
+        return rankings.size();
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/IndexResultValuator.java

@@ -7,6 +7,8 @@ import it.unimi.dsi.fastutil.ints.IntArrayList;
 import it.unimi.dsi.fastutil.longs.Long2LongOpenHashMap;
 import nu.marginalia.wmsa.edge.index.model.EdgePageWordFlags;
 import nu.marginalia.wmsa.edge.index.model.EdgePageWordMetadata;
+import nu.marginalia.wmsa.edge.index.model.QueryStrategy;
+import nu.marginalia.wmsa.edge.index.query.IndexQueryParams;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchResultItem;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchResultKeywordScore;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchSubquery;
@@ -17,6 +19,7 @@ import java.util.Objects;
 public class IndexResultValuator {
     private final IndexMetadataService metadataService;
     private final List<List<String>> searchTermVariants;
+    private final IndexQueryParams queryParams;
     private final int[] termIdsAll;
 
     private final TLongHashSet resultsWithPriorityTerms;
@@ -24,9 +27,10 @@ public class IndexResultValuator {
     private final TObjectIntHashMap<String> termToId = new TObjectIntHashMap<>(10, 0.75f, -1);
     private final TermMetadata termMetadata;
 
-    public IndexResultValuator(SearchIndexControl indexes, TLongList results, List<EdgeSearchSubquery> subqueries) {
+    public IndexResultValuator(SearchIndexControl indexes, TLongList results, List<EdgeSearchSubquery> subqueries, IndexQueryParams queryParams) {
         this.metadataService = new IndexMetadataService(indexes);
         this.searchTermVariants = subqueries.stream().map(sq -> sq.searchTermsInclude).distinct().toList();
+        this.queryParams = queryParams;
 
         var lexiconReader = Objects.requireNonNull(indexes.getLexiconReader());
         IntArrayList termIdsList = new IntArrayList();
@@ -114,10 +118,15 @@ public class IndexResultValuator {
                     docMetadata,
                     resultsWithPriorityTerms.contains(searchResult.combinedId)
             );
+
             searchResult.scores.add(score);
 
             setScore += score.termValue();
 
+            if (!filterRequired(metadata, queryParams.queryStrategy())) {
+                setScore += 1000;
+            }
+
             if (termIdx == 0) {
                 setScore += score.documentValue();
             }
@@ -130,6 +139,19 @@ public class IndexResultValuator {
         return setScore/setSize;
     }
 
+    private boolean filterRequired(long metadata, QueryStrategy queryStrategy) {
+        if (queryStrategy == QueryStrategy.REQUIRE_FIELD_SITE) {
+            return EdgePageWordFlags.Site.isPresent(metadata);
+        }
+        else if (queryStrategy == QueryStrategy.REQUIRE_FIELD_SUBJECT) {
+            return EdgePageWordFlags.Subjects.isPresent(metadata);
+        }
+        else if (queryStrategy == QueryStrategy.REQUIRE_FIELD_TITLE) {
+            return EdgePageWordFlags.Title.isPresent(metadata);
+        }
+        return true;
+    }
+
     private double calculateTermCoherencePenalty(int urlId, TObjectIntHashMap<String> termToId, List<String> termList) {
         long maskDirectGenerous = ~0;
         long maskDirectRaw = ~0;
@@ -139,6 +161,9 @@ public class IndexResultValuator {
                               | EdgePageWordFlags.Subjects.asBit()
                               | EdgePageWordFlags.Synthetic.asBit();
 
+        int termCount = 0;
+        double tfIdfSum = 1.;
+
         for (String term : termList) {
             var meta = termMetadata.getTermMetadata(termToId.get(term), urlId);
             long positions;
@@ -156,18 +181,22 @@ public class IndexResultValuator {
                 maskDirectGenerous &= positions;
             }
 
+            termCount++;
+            tfIdfSum += EdgePageWordMetadata.decodeTfidf(meta);
         }
 
+        double avgTfIdf = termCount / tfIdfSum;
+
         if (maskAdjacent == 0) {
-            return 40;
+            return Math.max(-2, 40 - 0.5 * avgTfIdf);
         }
 
         if (maskDirectGenerous == 0) {
-            return 20;
+            return Math.max(-1, 20 - 0.3 *  avgTfIdf);
         }
 
         if (maskDirectRaw == 0) {
-            return 2;
+            return Math.max(-1, 15 - 0.2 *  avgTfIdf);
         }
 
         return Long.numberOfTrailingZeros(maskDirectGenerous)/5. - Long.bitCount(maskDirectGenerous);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/SearchIndex.java

@@ -92,7 +92,8 @@ public class SearchIndex {
         SearchIndexReader.IndexQueryBuilder query =
             switch(params.queryStrategy()) {
                 case SENTENCE               -> indexReader.findWordAsSentence(orderedIncludes);
-                case TOPIC                  -> indexReader.findWordAsTopic(orderedIncludes);
+                case TOPIC, REQUIRE_FIELD_SITE, REQUIRE_FIELD_TITLE, REQUIRE_FIELD_SUBJECT
+                                            -> indexReader.findWordAsTopic(orderedIncludes);
                 case AUTO                   -> indexReader.findWordTopicDynamicMode(orderedIncludes);
             };
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/SearchIndexControl.java

@@ -6,6 +6,7 @@ import nu.marginalia.wmsa.configuration.server.Initialization;
 import nu.marginalia.wmsa.edge.index.IndexServicesFactory;
 import nu.marginalia.wmsa.edge.index.lexicon.KeywordLexiconReadOnlyView;
 import nu.marginalia.wmsa.edge.index.postings.journal.writer.SearchIndexJournalWriterImpl;
+import nu.marginalia.wmsa.edge.index.svc.EdgeIndexSearchSetsService;
 import nu.marginalia.wmsa.edge.index.svc.EdgeOpsLockService;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
@@ -26,13 +27,14 @@ public class SearchIndexControl {
 
     @Inject
     public SearchIndexControl(IndexServicesFactory servicesFactory,
-                              EdgeOpsLockService opsLockService) {
+                              EdgeOpsLockService opsLockService,
+                              EdgeIndexSearchSetsService searchSetsService) {
         this.servicesFactory = servicesFactory;
 
         this.primaryIndexWriter = servicesFactory.getIndexWriter(0);
         this.secondaryIndexWriter = servicesFactory.getIndexWriter(1);
 
-        index = servicesFactory.createIndexBucket();
+        index = servicesFactory.createIndexBucket(searchSetsService);
         this.opsLockService = opsLockService;
     }
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/forward/ForwardIndexConverter.java

@@ -3,6 +3,8 @@ package nu.marginalia.wmsa.edge.index.postings.forward;
 import com.upserve.uppend.blobs.NativeIO;
 import it.unimi.dsi.fastutil.longs.Long2IntOpenHashMap;
 import nu.marginalia.util.array.LongArray;
+import nu.marginalia.wmsa.edge.index.model.EdgePageDocumentsMetadata;
+import nu.marginalia.wmsa.edge.index.postings.DomainRankings;
 import nu.marginalia.wmsa.edge.index.postings.journal.reader.SearchIndexJournalReader;
 import nu.marginalia.wmsa.edge.index.postings.journal.reader.SearchIndexJournalReaderSingleFile;
 import org.roaringbitmap.IntConsumer;
@@ -18,26 +20,26 @@ import java.nio.file.Path;
 import static nu.marginalia.wmsa.edge.index.postings.forward.ForwardIndexParameters.*;
 
 public class ForwardIndexConverter {
-    private static final int RWF_BIN_SIZE = 10_000_000;
 
-    private final Path tmpFileDir;
     private final File inputFile;
 
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
     private final Path outputFileDocsId;
     private final Path outputFileDocsData;
+    private final DomainRankings domainRankings;
 
 
-    public ForwardIndexConverter(Path tmpFileDir,
+    public ForwardIndexConverter(
                                  File inputFile,
                                  Path outputFileDocsId,
-                                 Path outputFileDocsData
+                                 Path outputFileDocsData,
+                                 DomainRankings domainRankings
                                  ) {
-        this.tmpFileDir = tmpFileDir;
         this.inputFile = inputFile;
         this.outputFileDocsId = outputFileDocsId;
         this.outputFileDocsData = outputFileDocsData;
+        this.domainRankings = domainRankings;
     }
 
     public void convert() throws IOException {
@@ -50,6 +52,8 @@ public class ForwardIndexConverter {
 
         logger.info("Converting  {} {}",inputFile, journalReader.fileHeader);
 
+        logger.info("Domain Rankings size = {}", domainRankings.size());
+
         try {
             LongArray docsFileId = getDocIds(outputFileDocsId, journalReader);
 
@@ -68,7 +72,10 @@ public class ForwardIndexConverter {
             journalReader.forEach(entry -> {
                 long entryOffset = (long) ENTRY_SIZE * docIdToIdx.get(entry.urlId());
 
-                docFileData.set(entryOffset + METADATA_OFFSET, entry.docMeta());
+                int ranking = domainRankings.getRanking(entry.domainId());
+                long meta = EdgePageDocumentsMetadata.encodeRank(entry.docMeta(), ranking);
+
+                docFileData.set(entryOffset + METADATA_OFFSET, meta);
                 docFileData.set(entryOffset + DOMAIN_OFFSET, entry.domainId());
             });
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/forward/ForwardIndexReader.java

@@ -29,20 +29,30 @@ public class ForwardIndexReader {
 
         logger.info("Switching forward index");
 
+        ids = loadIds(idsFile);
+        data = loadData(dataFile);
+    }
+
+    private static TLongIntHashMap loadIds(Path idsFile) throws IOException {
         var idsArray = LongArray.mmapRead(idsFile);
         idsArray.advice(NativeIO.Advice.Sequential);
 
-        ids = new TLongIntHashMap((int) idsArray.size(), 0.5f, -1, -1);
+        var ids = new TLongIntHashMap((int) idsArray.size(), 0.5f, -1, -1);
 
         // This hash table should be of the same size as the number of documents, so typically less than 1 Gb
         idsArray.forEach(0, idsArray.size(), (pos, val) -> {
             ids.put(val, (int) pos);
         });
 
-        data = LongArray.mmapRead(dataFile);
+        return ids;
+    }
 
+    private static LongArray loadData(Path dataFile) throws IOException {
+        var data = LongArray.mmapRead(dataFile);
 
         data.advice(NativeIO.Advice.Random);
+
+        return data;
     }
 
     private int idxForDoc(long docId) {
@@ -55,6 +65,7 @@ public class ForwardIndexReader {
 
         return data.get(ENTRY_SIZE * offset + METADATA_OFFSET);
     }
+
     public int getDomainId(long docId) {
         long offset = idxForDoc(docId);
         if (offset < 0) return 0;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/forward/ParamMatchingQueryFilter.java

@@ -16,7 +16,7 @@ public class ParamMatchingQueryFilter implements QueryFilterStepIf {
 
     @Override
     public boolean test(long docId) {
-        var post = forwardIndexReader.docPost(docId);
+        var post = forwardIndexReader.docPost(docId & 0xFFFF_FFFFL);
 
         if (!validateDomain(post)) {
             return false;
@@ -33,6 +33,11 @@ public class ParamMatchingQueryFilter implements QueryFilterStepIf {
         if (!validateSize(post)) {
             return false;
         }
+
+        if (!validateRank(post)) {
+            return false;
+        }
+
         return true;
     }
 
@@ -51,6 +56,7 @@ public class ParamMatchingQueryFilter implements QueryFilterStepIf {
 
         return limit.test(quality);
     }
+
     private boolean validateYear(ForwardIndexReader.DocPost post) {
         if (params.year().type() == SpecificationLimitType.NONE)
             return true;
@@ -69,6 +75,15 @@ public class ParamMatchingQueryFilter implements QueryFilterStepIf {
         return params.size().test(postVal);
     }
 
+    private boolean validateRank(ForwardIndexReader.DocPost post) {
+        if (params.rank().type() == SpecificationLimitType.NONE)
+            return true;
+
+        int postVal = EdgePageDocumentsMetadata.decodeRank(post.meta());
+
+        return params.rank().test(postVal);
+    }
+
     @Override
     public double cost() {
         return 32;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/reverse/ReverseIndexConverter.java

@@ -8,6 +8,7 @@ import nu.marginalia.util.array.functional.LongBinaryIOOperation;
 import nu.marginalia.util.array.functional.LongIOTransformer;
 import nu.marginalia.util.array.functional.LongTransformer;
 import nu.marginalia.util.btree.BTreeWriter;
+import nu.marginalia.wmsa.edge.index.postings.DomainRankings;
 import nu.marginalia.wmsa.edge.index.postings.journal.model.SearchIndexJournalEntry;
 import nu.marginalia.wmsa.edge.index.postings.journal.model.SearchIndexJournalStatistics;
 import nu.marginalia.wmsa.edge.index.postings.journal.reader.SearchIndexJournalReader;
@@ -32,18 +33,22 @@ public class ReverseIndexConverter {
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
     private final SearchIndexJournalReaderSingleFile journalReader;
+    private final DomainRankings domainRankings;
     private final Path outputFileWords;
     private final Path outputFileDocs;
-
+    private final SortingContext sortingContext;
 
     public ReverseIndexConverter(Path tmpFileDir,
                                  SearchIndexJournalReaderSingleFile journalReader,
+                                 DomainRankings domainRankings,
                                  Path outputFileWords,
                                  Path outputFileDocs) {
         this.tmpFileDir = tmpFileDir;
         this.journalReader = journalReader;
+        this.domainRankings = domainRankings;
         this.outputFileWords = outputFileWords;
         this.outputFileDocs = outputFileDocs;
+        this.sortingContext = new SortingContext(tmpFileDir, 64_000);
     }
 
     public void convert() throws IOException {
@@ -56,7 +61,7 @@ public class ReverseIndexConverter {
         final SearchIndexJournalStatistics statistics = journalReader.getStatistics();
 
         final Path intermediateUrlsFile = Files.createTempFile(tmpFileDir, "urls-sorted", ".dat");
-        SortingContext sortingContext = new SortingContext(tmpFileDir, 64_000);
+
 
         try {
             final long wordsFileSize = statistics.highestWord() + 1;
@@ -187,7 +192,7 @@ public class ReverseIndexConverter {
         }
     }
 
-    private static class IntermediateIndexConstructor implements SearchIndexJournalReaderSingleFile.LongObjectConsumer<SearchIndexJournalEntry.Record>, AutoCloseable {
+    private class IntermediateIndexConstructor implements SearchIndexJournalReaderSingleFile.LongObjectConsumer<SearchIndexJournalEntry.Record>, AutoCloseable {
 
         private final LongArray wordRangeEnds;
         private final IntArray wordRangeOffset;
@@ -205,12 +210,26 @@ public class ReverseIndexConverter {
 
         @Override
         public void accept(long docId, SearchIndexJournalEntry.Record record) {
-            final long urlId = docId & 0xFFFF_FFFFL;
-            final int wordId = record.wordId();
 
+            /* Encode the ID as
+             *
+             *     32 bits  32 bits
+             *   [ ranking | url-id ]
+             *
+             *  in order to get low-ranking documents to be considered first
+             *  when sorting the items.
+             */
+
+            int domainId = (int) (docId >>> 32);
+            long rankingId = (long) domainRankings.getRanking(domainId) << 32;
+
+            int urlId = (int) (docId & 0xFFFF_FFFFL);
+            long rankEncodedId = rankingId | urlId;
+
+            final int wordId = record.wordId();
             long offset = startOfRange(wordId);
 
-            documentsFile.put(offset + wordRangeOffset.getAndIncrement(wordId), urlId);
+            documentsFile.put(offset + wordRangeOffset.getAndIncrement(wordId), rankEncodedId);
             documentsFile.put(offset + wordRangeOffset.getAndIncrement(wordId), record.metadata());
 
         }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/reverse/ReverseIndexPrioReader.java

@@ -47,18 +47,6 @@ public class ReverseIndexPrioReader {
         return new ReverseIndexEntrySource(createReaderNew(offset), ReverseIndexEntrySourceBehavior.DO_PREFER);
     }
 
-    public int numDocuments(int wordId) {
-        if (wordId < 0)
-            return 0;
-
-        long offset = words.get(wordId);
-
-        if (offset < 0)
-            return 0;
-
-        return createReaderNew(offset).numEntries();
-    }
-
     private BTreeReader createReaderNew(long offset) {
         return new BTreeReader(documents, ReverseIndexParameters.bTreeContext, offset);
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/postings/reverse/ReverseIndexReader.java

@@ -53,6 +53,11 @@ public class ReverseIndexReader {
     }
 
     public EntrySource documents(int wordId, ReverseIndexEntrySourceBehavior behavior) {
+        if (null == words) {
+            logger.warn("Reverse index is not ready, dropping query");
+            return new EmptyEntrySource();
+        }
+
         if (wordId < 0 || wordId >= words.size()) return new EmptyEntrySource();
 
         long offset = words.get(wordId);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/query/IndexQueryParams.java

@@ -7,6 +7,7 @@ import nu.marginalia.wmsa.edge.model.search.domain.SpecificationLimit;
 public record IndexQueryParams(SpecificationLimit qualityLimit,
                                SpecificationLimit year,
                                SpecificationLimit size,
+                               SpecificationLimit rank,
                                SearchSet searchSet,
                                QueryStrategy queryStrategy
                                )

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/RankingAlgorithm.java

@@ -1,21 +1,21 @@
-package nu.marginalia.util.ranking;
+package nu.marginalia.wmsa.edge.index.ranking;
 
-import gnu.trove.list.TIntList;
 import gnu.trove.list.array.TIntArrayList;
 import gnu.trove.map.hash.TIntIntHashMap;
 import gnu.trove.map.hash.TIntObjectHashMap;
 import it.unimi.dsi.fastutil.ints.IntArrays;
-import it.unimi.dsi.fastutil.ints.IntComparator;
-import org.roaringbitmap.RoaringBitmap;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultAccumulator;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainData;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
 import java.util.Arrays;
-import java.util.Comparator;
 import java.util.HashSet;
 import java.util.Set;
-import java.util.function.IntToDoubleFunction;
-import java.util.stream.IntStream;
+import java.util.function.Supplier;
+
+import static java.lang.Math.min;
 
 public abstract class RankingAlgorithm {
     protected final TIntObjectHashMap<RankingDomainData> domainsById = new TIntObjectHashMap<>();
@@ -85,6 +85,10 @@ public abstract class RankingAlgorithm {
         logger.info("Origin Domains: {}", originDomainIds.size());
     }
 
+    public RankingDomainData getDomainData(int id) {
+        return domainsById.get(id);
+    }
+
     public void addPeripheralNodes() {
 
         int newNodesIdxCutoff = domainIdToIndex.size();
@@ -133,29 +137,7 @@ public abstract class RankingAlgorithm {
         return domainsById.size();
     }
 
-
-    public RankVector pageRankVector() {
-        RankVector rank = new RankVector(1.d / domainsById.size());
-
-        int iter_max = 100;
-        for (int i = 0; i < iter_max; i++) {
-            RankVector newRank = createNewRankVector(rank);
-
-            double oldNorm = rank.norm();
-            double newNorm = newRank.norm();
-            double dNorm = oldNorm - newNorm ;
-            if (i < iter_max-1) {
-                adjustRankVector(newRank, dNorm, oldNorm);
-            }
-
-            rank = newRank;
-        }
-
-        return rank;
-    }
-
-
-    public RoaringBitmap pageRank(int resultCount) {
+    public <T> T pageRank(int resultCount, Supplier<RankingResultAccumulator<T>> accumulatorP) {
         RankVector rank = new RankVector(1.d / domainsById.size());
 
         int iter_max = 100;
@@ -174,10 +156,10 @@ public abstract class RankingAlgorithm {
         }
 
 
-        return rank.getRanking(resultCount);
+        return rank.getRanking(resultCount, accumulatorP).get();
     }
 
-    public RoaringBitmap pageRankWithPeripheralNodes(int resultCount) {
+    public <T> T pageRankWithPeripheralNodes(int resultCount, Supplier<RankingResultAccumulator<T>> accumulatorP) {
         RankVector rank = new RankVector(1.d / domainsById.size());
 
         int iter_max = 100;
@@ -201,32 +183,11 @@ public abstract class RankingAlgorithm {
 
         logger.info("PRWPN iteration done");
 
-        return rank.getRanking(resultCount);
+        return rank.getRanking(resultCount, accumulatorP).get();
     }
 
     abstract void adjustRankVector(RankVector vector, double dNorm, double oldNorm);
 
-    public TIntList pageRank(IntToDoubleFunction weight, int resultCount) {
-        RankVector rank = new RankVector(1.d / domainsById.size());
-
-        int iter_max = 100;
-        for (int i = 0; i < iter_max; i++) {
-            RankVector newRank = createNewRankVector(rank);
-
-            double oldNorm = rank.norm();
-            double newNorm = newRank.norm();
-            double dNorm = oldNorm - newNorm ;
-
-            if (i < iter_max-1) {
-                adjustRankVector(newRank, dNorm, oldNorm);
-            }
-
-            rank = newRank;
-        }
-
-        return rank.getRanking(weight, resultCount);
-    }
-
     abstract RankVector createNewRankVector(RankVector rank);
 
     public boolean includeInRanking(RankingDomainData data) {
@@ -245,9 +206,9 @@ public abstract class RankingAlgorithm {
     public void setMaxKnownUrls(int maxKnownUrls) {
         this.maxKnownUrls = maxKnownUrls;
     }
-
     public class RankVector {
         private final double[] rank;
+
         public RankVector(double defaultValue) {
             rank = new double[domainIndexToId.size()];
             if (defaultValue != 0.) {
@@ -271,9 +232,8 @@ public abstract class RankingAlgorithm {
 
         public double norm() {
             double v = 0.;
-            for (int i = 0; i < rank.length; i++) {
-                if (rank[i] > 0) { v+=rank[i]; }
-                else { v -= rank[i]; }
+            for (double value : rank) {
+                v += Math.abs(value);
             }
             return v;
         }
@@ -281,74 +241,39 @@ public abstract class RankingAlgorithm {
         public double norm(RankVector other) {
             double v = 0.;
             for (int i = 0; i < rank.length; i++) {
-                double dv = rank[i] - other.get(i);
-
-                if (dv > 0) { v+=dv; }
-                else { v -= dv; }
+                v += Math.abs(rank[i] - other.get(i));
             }
             return v;
         }
 
-        public TIntList getRanking(IntToDoubleFunction other, int numResults) {
-            TIntArrayList list = new TIntArrayList(numResults);
+        public <T> RankingResultAccumulator<T> getRanking(int numResults, Supplier<RankingResultAccumulator<T>> accumulatorP) {
 
-            Comparator<Integer> comparator = Comparator.comparing(i -> Math.sqrt(other.applyAsDouble(domainIdToIndex.get(i)) * rank[i]));
-
-            IntStream.range(0, rank.length)
-                    .boxed()
-                    .sorted(comparator.reversed())
-                    .map(domainIndexToId::get)
-                    .limit(numResults)
-                    .forEach(list::add);
-
-            return list;
-        }
-
-        public RoaringBitmap getRanking(int numResults) {
             if (numResults < 0) {
                 numResults = domainIdToIndex.size();
             }
-            if (numResults >= rank.length) {
-                numResults = rank.length;
-            }
+            numResults = min(numResults, min(domainIdToIndex.size(), rank.length));
 
-            RoaringBitmap list = new RoaringBitmap();
+            int[] nodes = sortOrder(rank);
+            var accumulator = accumulatorP.get();
 
-            int[] nodes = new int[rank.length];
-            Arrays.setAll(nodes, i->i);
-            IntComparator comp = (i,j) -> (int) Math.signum(rank[j] - rank[i]);
-            IntArrays.quickSort(nodes, comp);
-
-            int i;
-
-            for (i = 0; i < numResults; i++) {
+            for (int i = 0; i < numResults; i++) {
                 int id = domainIndexToId.get(nodes[i]);
 
                 if (includeInRanking(domainsById.get(id)))
-                    list.add(id);
+                    accumulator.add(id, i);
             }
 
-            for (; i < nodes.length && domainsById.size() < numResults; i++) {
-                int id = domainIndexToId.get(nodes[i]);
+            return accumulator;
+        }
+        private static int[] sortOrder(double[] values) {
 
-                if (includeInRanking(domainsById.get(id)))
-                    list.add(id);
-            }
+            int[] ret = new int[values.length];
+            Arrays.setAll(ret, i->i);
+            IntArrays.quickSort(ret, (i,j) -> (int) Math.signum(values[j] - values[i]));
 
-
-            return list;
+            return ret;
         }
 
-
-        public void incrementAll(double v) {
-            for (int i = 0; i < rank.length; i++) {
-                rank[i]+=v;
-            }
-        }
-
-        int size() {
-            return domainsById.size();
-        }
     }
 
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/ReversePageRank.java

@@ -1,10 +1,12 @@
-package nu.marginalia.util.ranking;
+package nu.marginalia.wmsa.edge.index.ranking;
 
 
-public class BetterReversePageRank extends RankingAlgorithm {
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
+
+public class ReversePageRank extends RankingAlgorithm {
 
 
-    public BetterReversePageRank(RankingDomainFetcher domains, String... origins) {
+    public ReversePageRank(RankingDomainFetcher domains, String... origins) {
         super(domains, origins);
     }
 
@@ -20,8 +22,6 @@ public class BetterReversePageRank extends RankingAlgorithm {
             double newRankValue = 0;
 
             if (links != null && links.size() > 0) {
-
-
                 for (int j = 0; j < links.size(); j++) {
                     var revLinks = linkDataDest2Src[links.getQuick(j)];
                     newRankValue += rank.get(links.getQuick(j)) / revLinks.size();

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/StandardPageRank.java

@@ -1,9 +1,11 @@
-package nu.marginalia.util.ranking;
+package nu.marginalia.wmsa.edge.index.ranking;
 
 
-public class BetterStandardPageRank extends RankingAlgorithm {
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
 
-    public BetterStandardPageRank(RankingDomainFetcher domains, String... origins) {
+public class StandardPageRank extends RankingAlgorithm {
+
+    public StandardPageRank(RankingDomainFetcher domains, String... origins) {
         super(domains, origins);
     }
 
@@ -38,8 +40,7 @@ public class BetterStandardPageRank extends RankingAlgorithm {
 
     @Override
     void adjustRankVector(RankVector vector, double dNorm, double oldNorm) {
-        originDomainIds.forEach(id -> vector.increment(id, 0.15 / originDomainIds.size() /* dNorm/originDomainIds.size() */ ));
-//        vector.incrementAll(0.14*dNorm/vector.size());
+        originDomainIds.forEach(id -> vector.increment(id, 0.15 / originDomainIds.size() ));
     }
 
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/accumulator/RankingResultAccumulator.java

@@ -0,0 +1,6 @@
+package nu.marginalia.wmsa.edge.index.ranking.accumulator;
+
+public interface RankingResultAccumulator<T> {
+    void add(int domainId, int rank);
+    T get();
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/accumulator/RankingResultBitSetAccumulator.java

@@ -0,0 +1,17 @@
+package nu.marginalia.wmsa.edge.index.ranking.accumulator;
+
+import org.roaringbitmap.RoaringBitmap;
+
+public class RankingResultBitSetAccumulator implements RankingResultAccumulator<RoaringBitmap> {
+    private final RoaringBitmap result = new RoaringBitmap();
+
+    @Override
+    public void add(int domainId, int rank) {
+        result.add(domainId);
+    }
+
+    @Override
+    public RoaringBitmap get() {
+        return result;
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/accumulator/RankingResultHashMapAccumulator.java

@@ -0,0 +1,21 @@
+package nu.marginalia.wmsa.edge.index.ranking.accumulator;
+
+import it.unimi.dsi.fastutil.ints.Int2IntOpenHashMap;
+
+public class RankingResultHashMapAccumulator implements RankingResultAccumulator<Int2IntOpenHashMap> {
+    private final Int2IntOpenHashMap result;
+
+    public RankingResultHashMapAccumulator(int size) {
+         result = new Int2IntOpenHashMap(size);
+    }
+
+    @Override
+    public void add(int domainId, int rank) {
+        result.put(domainId, rank);
+    }
+
+    @Override
+    public Int2IntOpenHashMap get() {
+        return result;
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/accumulator/RankingResultListAccumulator.java

@@ -0,0 +1,24 @@
+package nu.marginalia.wmsa.edge.index.ranking.accumulator;
+
+import gnu.trove.list.array.TIntArrayList;
+
+public class RankingResultListAccumulator implements RankingResultAccumulator<TIntArrayList> {
+    private final TIntArrayList result;
+
+    public RankingResultListAccumulator(int size) {
+         result = new TIntArrayList(size);
+    }
+    public RankingResultListAccumulator() {
+         result = new TIntArrayList(10_000);
+    }
+
+    @Override
+    public void add(int domainId, int rank) {
+        result.add(domainId);
+    }
+
+    @Override
+    public TIntArrayList get() {
+        return result;
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/data/RankingDomainData.java

@@ -1,4 +1,4 @@
-package nu.marginalia.util.ranking;
+package nu.marginalia.wmsa.edge.index.ranking.data;
 
 import lombok.AllArgsConstructor;
 import lombok.Data;
@@ -10,7 +10,7 @@ public class RankingDomainData {
     public final int id;
     public final String name;
     private int alias;
-    private EdgeDomainIndexingState state;
+    public EdgeDomainIndexingState state;
     public final int knownUrls;
 
     public int resolveAlias() {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/data/RankingDomainFetcher.java

@@ -1,6 +1,7 @@
-package nu.marginalia.util.ranking;
+package nu.marginalia.wmsa.edge.index.ranking.data;
 
 import com.google.inject.Inject;
+import com.google.inject.Singleton;
 import com.zaxxer.hikari.HikariDataSource;
 import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
 import nu.marginalia.wmsa.edge.model.crawl.EdgeDomainIndexingState;
@@ -11,12 +12,13 @@ import java.sql.SQLException;
 import java.util.function.Consumer;
 import java.util.function.IntConsumer;
 
+@Singleton
 public class RankingDomainFetcher {
-    private final HikariDataSource dataSource;
-    private final EdgeDomainBlacklistImpl blacklist;
-    private final Logger logger = LoggerFactory.getLogger(getClass());
+    protected final HikariDataSource dataSource;
+    protected final EdgeDomainBlacklistImpl blacklist;
+    protected final Logger logger = LoggerFactory.getLogger(getClass());
 
-    private final boolean getNames = false;
+    protected boolean getNames = false;
 
     @Inject
     public RankingDomainFetcher(HikariDataSource dataSource, EdgeDomainBlacklistImpl blacklist) {
@@ -24,6 +26,10 @@ public class RankingDomainFetcher {
         this.blacklist = blacklist;
     }
 
+    public void retainNames() {
+        this.getNames = true;
+    }
+
     public void getDomains(Consumer<RankingDomainData> consumer) {
         String query;
         if (getNames) {
@@ -49,14 +55,19 @@ public class RankingDomainFetcher {
         getDomains(query, consumer);
     }
 
-    private void getDomains(String query, Consumer<RankingDomainData> consumer) {
+    protected void getDomains(String query, Consumer<RankingDomainData> consumer) {
         try (var conn = dataSource.getConnection(); var stmt = conn.prepareStatement(query)) {
             stmt.setFetchSize(10000);
             var rsp = stmt.executeQuery();
             while (rsp.next()) {
                 int id = rsp.getInt(1);
                 if (!blacklist.isBlacklisted(id)) {
-                    consumer.accept(new RankingDomainData(id, rsp.getString(2), rsp.getInt(3), EdgeDomainIndexingState.valueOf(rsp.getString(4)), rsp.getInt(5)));
+                    consumer.accept(
+                            new RankingDomainData(id,
+                                    rsp.getString(2),
+                                    rsp.getInt(3),
+                                    EdgeDomainIndexingState.valueOf(rsp.getString(4)),
+                                    rsp.getInt(5)));
                 }
             }
         }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/data/RankingDomainFetcherForSimilarityData.java

@@ -0,0 +1,103 @@
+package nu.marginalia.wmsa.edge.index.ranking.data;
+
+import com.google.inject.Inject;
+import com.google.inject.Singleton;
+import com.zaxxer.hikari.HikariDataSource;
+import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
+import org.slf4j.LoggerFactory;
+
+import java.sql.SQLException;
+import java.util.function.Consumer;
+
+@Singleton
+public class RankingDomainFetcherForSimilarityData extends RankingDomainFetcher {
+    final boolean hasData;
+
+    @Inject
+    public RankingDomainFetcherForSimilarityData(HikariDataSource dataSource, EdgeDomainBlacklistImpl blacklist) {
+        super(dataSource, blacklist);
+
+        hasData = isDomainNeighborTablePopulated(dataSource);
+    }
+
+    private static boolean isDomainNeighborTablePopulated(HikariDataSource dataSource) {
+        try (var conn = dataSource.getConnection();
+             var stmt = conn.createStatement();
+             var rs = stmt.executeQuery("SELECT DOMAIN_ID FROM EC_DOMAIN_NEIGHBORS_2 LIMIT 1")) {
+
+            return rs.next();
+        }
+        catch (SQLException ex) {
+            LoggerFactory
+                    .getLogger(RankingDomainFetcherForSimilarityData.class)
+                    .error("Failed to get count from EC_DOMAIN_NEIGHBORS_2", ex);
+            return false;
+        }
+    }
+    public boolean hasData() {
+        return hasData;
+    }
+
+    public void eachDomainLink(DomainLinkConsumer consumer) {
+        try (var conn = dataSource.getConnection();
+             var stmt = conn.prepareStatement("SELECT DOMAIN_ID, NEIGHBOR_ID, RELATEDNESS FROM EC_DOMAIN_NEIGHBORS_2"))
+        {
+            stmt.setFetchSize(10000);
+
+            var rsp = stmt.executeQuery();
+
+            while (rsp.next()) {
+                int src = rsp.getInt(1);
+                int dst = rsp.getInt(2);
+
+                // these "links" are bidi
+                consumer.accept(src, dst);
+                consumer.accept(dst, src);
+            }
+        }
+        catch (SQLException ex) {
+            logger.error("Failed to fetch domain links", ex);
+        }
+    }
+
+    public void getDomains(Consumer<RankingDomainData> consumer) {
+//        String query =
+//               """
+//                   SELECT EC_DOMAIN.ID,DOMAIN_NAME,DOMAIN_ALIAS,STATE,COALESCE(KNOWN_URLS, 0)
+//                   FROM EC_DOMAIN
+//                   LEFT JOIN DOMAIN_METADATA ON EC_DOMAIN.ID=DOMAIN_METADATA.ID
+//                   INNER JOIN EC_DOMAIN_LINK ON DEST_DOMAIN_ID=EC_DOMAIN.ID
+//                   WHERE SOURCE_DOMAIN_ID!=DEST_DOMAIN_ID
+//                   GROUP BY EC_DOMAIN.ID
+//                   HAVING COUNT(SOURCE_DOMAIN_ID)>5
+//               """;
+
+        String query;
+        if (getNames) {
+            query =
+                    """
+                        SELECT EC_DOMAIN.ID,DOMAIN_NAME,DOMAIN_ALIAS,STATE,COALESCE(KNOWN_URLS, 0)
+                        FROM EC_DOMAIN 
+                        INNER JOIN DOMAIN_METADATA ON EC_DOMAIN.ID=DOMAIN_METADATA.ID 
+                        GROUP BY EC_DOMAIN.ID 
+                    """;
+        }
+        else {
+            query =
+                    """
+                        SELECT EC_DOMAIN.ID,"",DOMAIN_ALIAS,STATE,COALESCE(KNOWN_URLS, 0)
+                        FROM EC_DOMAIN 
+                        INNER JOIN DOMAIN_METADATA ON EC_DOMAIN.ID=DOMAIN_METADATA.ID 
+                        GROUP BY EC_DOMAIN.ID 
+                    """;
+        }
+
+        getDomains(query, consumer);
+    }
+
+
+    public void getPeripheralDomains(Consumer<RankingDomainData> consumer) {
+        // This is not relevant for this variant of pagerank since it is bidirectional
+    }
+
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/old/OldReversePageRankV2.java

@@ -1,4 +1,4 @@
-package nu.marginalia.util.ranking.old;
+package nu.marginalia.wmsa.edge.index.ranking.old;
 
 
 import com.zaxxer.hikari.HikariDataSource;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/old/StandardPageRank.java

@@ -1,4 +1,4 @@
-package nu.marginalia.util.ranking.old;
+package nu.marginalia.wmsa.edge.index.ranking.old;
 
 
 import com.zaxxer.hikari.HikariDataSource;
@@ -125,7 +125,6 @@ public class StandardPageRank {
 
         final TIntArrayList empty = new TIntArrayList();
 
-        double rankNorm = rank.norm();
         RankVector newRank = new RankVector(0);
 
         for (DomainData domain : domains.valueCollection()) {
@@ -176,8 +175,6 @@ public class StandardPageRank {
                 }
             });
         }
-
-        TIntHashSet deadEnds = new TIntHashSet(domains.size());
     }
 
     private class RankVector {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/tool/CreateBrowseDomainRanksTool.java

@@ -1,43 +1,30 @@
-package nu.marginalia.util.ranking.tool;
+package nu.marginalia.wmsa.edge.index.ranking.tool;
 
 import com.zaxxer.hikari.HikariDataSource;
 import lombok.SneakyThrows;
-import nu.marginalia.util.ranking.BuggyStandardPageRank;
-import nu.marginalia.util.ranking.RankingDomainFetcher;
 import nu.marginalia.wmsa.configuration.module.DatabaseModule;
 import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
+import nu.marginalia.wmsa.edge.index.ranking.StandardPageRank;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultListAccumulator;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcherForSimilarityData;
 import org.mariadb.jdbc.Driver;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
 import java.sql.SQLException;
-import java.util.HashSet;
-import java.util.Set;
 import java.util.concurrent.LinkedBlockingQueue;
 
-public class UpdateDomainRanksTool {
+public class CreateBrowseDomainRanksTool {
 
-    private static final Logger logger = LoggerFactory.getLogger(UpdateDomainRanksTool.class);
+    private static final Logger logger = LoggerFactory.getLogger(CreateBrowseDomainRanksTool.class);
 
-    public Set<String> originDomains = new HashSet<>();
-    public Set<Integer> originDomainIds = new HashSet<>();
-    public final long domainIdMax = -1;
-    public int domainCount;
-    private volatile static int rankMax;
-
-    public int maxId() {
-        return (int) domainIdMax;
-    }
-    public int domainCount() {
-        return domainCount;
-    }
 
     static final LinkedBlockingQueue<Integer> uploadQueue = new LinkedBlockingQueue<>(10);
     volatile static boolean running = true;
 
     @SneakyThrows
     public static void main(String... args) {
-        org.mariadb.jdbc.Driver driver = new Driver();
+        Driver driver = new Driver();
         var conn = new DatabaseModule().provideConnection();
 
         long start = System.currentTimeMillis();
@@ -45,20 +32,21 @@ public class UpdateDomainRanksTool {
 
         logger.info("Ranking");
         var ds = new DatabaseModule().provideConnection();
-        var domains = new RankingDomainFetcher(ds, new EdgeDomainBlacklistImpl(ds));
-        var spr = new BuggyStandardPageRank(domains, "memex.marginalia.nu");
+        var domains = new RankingDomainFetcherForSimilarityData(ds, new EdgeDomainBlacklistImpl(ds));
+        var rpr = new StandardPageRank(domains,  args);
 
-        rankMax = spr.size()*2;
         uploader.start();
 
-        var rankData = spr.pageRankWithPeripheralNodes(rankMax);
-        for (int i : rankData) {
+        var rankData = rpr.pageRankWithPeripheralNodes(1000, RankingResultListAccumulator::new);
+
+        rankData.forEach(i -> {
             try {
                 uploadQueue.put(i);
             } catch (InterruptedException e) {
                 e.printStackTrace();
             }
-        }
+            return true;
+        });
 
         long end = System.currentTimeMillis();
         running = false;
@@ -68,24 +56,14 @@ public class UpdateDomainRanksTool {
     }
 
     public static void uploadThread(HikariDataSource dataSource) {
-        int i = 0;
-
         try (var conn = dataSource.getConnection()) {
-            logger.info("Resetting rank");
-            try (var stmt = conn.prepareStatement("UPDATE EC_DOMAIN SET RANK=1")) {
-                stmt.executeUpdate();
-            }
-
-            logger.info("Updating ranks");
-            try (var stmt = conn.prepareStatement("UPDATE EC_DOMAIN SET RANK=? WHERE ID=?")) {
+            try (var stmt = conn.prepareStatement("INSERT IGNORE INTO EC_RANDOM_DOMAINS(DOMAIN_SET, DOMAIN_ID) VALUES (3, ?)")) {
                 while (running || (!running && !uploadQueue.isEmpty())) {
                     var job = uploadQueue.take();
-                    stmt.setDouble(1, i++ / (double) rankMax);
-                    stmt.setInt(2, job);
+                    stmt.setInt(1, job);
                     stmt.executeUpdate();
                 }
             }
-
         } catch (SQLException | InterruptedException throwables) {
             throwables.printStackTrace();
         }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/tool/PerusePageRankV2.java

@@ -1,4 +1,4 @@
-package nu.marginalia.util.ranking.tool;
+package nu.marginalia.wmsa.edge.index.ranking.tool;
 
 
 import com.zaxxer.hikari.HikariDataSource;
@@ -10,9 +10,9 @@ import it.unimi.dsi.fastutil.ints.IntArrays;
 import it.unimi.dsi.fastutil.ints.IntComparator;
 import lombok.AllArgsConstructor;
 import lombok.SneakyThrows;
-import nu.marginalia.util.ranking.RankingAlgorithm;
-import nu.marginalia.util.ranking.RankingDomainData;
-import nu.marginalia.util.ranking.RankingDomainFetcher;
+import nu.marginalia.wmsa.edge.index.ranking.RankingAlgorithm;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainData;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
 import nu.marginalia.wmsa.configuration.module.DatabaseModule;
 import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
 import org.jetbrains.annotations.NotNull;
@@ -33,8 +33,6 @@ public class PerusePageRankV2 {
     TIntArrayList[] linkDataSrc2Dest;
     TIntArrayList[] linkDataDest2Src;
 
-    private static final boolean getNames = true;
-
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
     static final LinkedBlockingQueue<LinkAdjacencies> uploadQueue = new LinkedBlockingQueue<>(10);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/tool/PrintDomainRanksTool.java

@@ -0,0 +1,67 @@
+package nu.marginalia.wmsa.edge.index.ranking.tool;
+
+import lombok.SneakyThrows;
+import nu.marginalia.wmsa.configuration.module.DatabaseModule;
+import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
+import nu.marginalia.wmsa.edge.index.ranking.StandardPageRank;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultListAccumulator;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcherForSimilarityData;
+import org.mariadb.jdbc.Driver;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.util.concurrent.LinkedBlockingQueue;
+import java.util.concurrent.atomic.AtomicInteger;
+
+public class PrintDomainRanksTool {
+
+    private static final Logger logger = LoggerFactory.getLogger(PrintDomainRanksTool.class);
+
+    private volatile static int rankMax;
+
+    static final LinkedBlockingQueue<Integer> uploadQueue = new LinkedBlockingQueue<>(10);
+    volatile static boolean running = true;
+
+    @SneakyThrows
+    public static void main(String... args) {
+        Driver driver = new Driver();
+        var conn = new DatabaseModule().provideConnection();
+
+        long start = System.currentTimeMillis();
+
+        logger.info("Ranking");
+        var ds = new DatabaseModule().provideConnection();
+
+        RankingDomainFetcher domains;
+        if (Boolean.getBoolean("use-link-data")) {
+            domains = new RankingDomainFetcher(ds, new EdgeDomainBlacklistImpl(ds));
+            domains.retainNames();
+        }
+        else {
+            domains = new RankingDomainFetcherForSimilarityData(ds, new EdgeDomainBlacklistImpl(ds));
+            domains.retainNames();
+        }
+
+        var rpr = new StandardPageRank(domains,  args);
+
+        rankMax = rpr.size();
+
+        var rankData = rpr.pageRankWithPeripheralNodes(rankMax, RankingResultListAccumulator::new);
+
+        AtomicInteger cnt = new AtomicInteger();
+        rankData.forEach(i -> {
+
+            var data = rpr.getDomainData(i);
+
+            System.out.printf("%d %s %s\n", cnt.getAndIncrement(), data.name, data.state);
+            return true;
+        });
+
+        long end = System.currentTimeMillis();
+        running = false;
+
+        logger.info("Done in {}", (end - start)/1000.0);
+    }
+
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/ranking/tool/UpdateDomainRanksTool.java

@@ -1,11 +1,12 @@
-package nu.marginalia.util.ranking.tool;
+package nu.marginalia.wmsa.edge.index.ranking.tool;
 
 import com.zaxxer.hikari.HikariDataSource;
 import lombok.SneakyThrows;
-import nu.marginalia.util.ranking.BetterReversePageRank;
-import nu.marginalia.util.ranking.RankingDomainFetcher;
+import nu.marginalia.wmsa.edge.index.ranking.StandardPageRank;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultListAccumulator;
 import nu.marginalia.wmsa.configuration.module.DatabaseModule;
 import nu.marginalia.wmsa.edge.dbcommon.EdgeDomainBlacklistImpl;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcherForSimilarityData;
 import org.mariadb.jdbc.Driver;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
@@ -13,12 +14,10 @@ import org.slf4j.LoggerFactory;
 import java.sql.SQLException;
 import java.util.concurrent.LinkedBlockingQueue;
 
-public class UpdateDomainRanksTool2 {
+public class UpdateDomainRanksTool {
 
-    private static final Logger logger = LoggerFactory.getLogger(UpdateDomainRanksTool2.class);
+    private static final Logger logger = LoggerFactory.getLogger(UpdateDomainRanksTool.class);
 
-    public final long domainIdMax = -1;
-    public int domainCount;
     private volatile static int rankMax;
 
     static final LinkedBlockingQueue<Integer> uploadQueue = new LinkedBlockingQueue<>(10);
@@ -34,21 +33,22 @@ public class UpdateDomainRanksTool2 {
 
         logger.info("Ranking");
         var ds = new DatabaseModule().provideConnection();
-        var domains = new RankingDomainFetcher(ds, new EdgeDomainBlacklistImpl(ds));
-         var rpr = new BetterReversePageRank(domains,  "memex.marginalia.nu", "bikobatanari.art", "sadgrl.online", "wiki.xxiivv.com", "%neocities.org");
+        var domains = new RankingDomainFetcherForSimilarityData(ds, new EdgeDomainBlacklistImpl(ds));
+        var rpr = new StandardPageRank(domains,  "memex.marginalia.nu", "bikobatanari.art", "sadgrl.online", "wiki.xxiivv.com");
 
-        var rankVector = rpr.pageRankVector();
         rankMax = rpr.size();
         uploader.start();
 
-        var rankData = rpr.pageRankWithPeripheralNodes(rankMax);
-        for (int i : rankData) {
+        var rankData = rpr.pageRankWithPeripheralNodes(rankMax, RankingResultListAccumulator::new);
+
+        rankData.forEach(i -> {
             try {
                 uploadQueue.put(i);
             } catch (InterruptedException e) {
                 e.printStackTrace();
             }
-        }
+            return true;
+        });
 
         long end = System.currentTimeMillis();
         running = false;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/EdgeIndexDomainQueryService.java

@@ -5,7 +5,7 @@ import com.google.inject.Inject;
 import com.google.inject.Singleton;
 import io.prometheus.client.Histogram;
 import nu.marginalia.util.array.buffer.LongQueryBuffer;
-import nu.marginalia.util.dict.DictionaryHashMap;
+import nu.marginalia.util.dict.OffHeapDictionaryHashMap;
 import nu.marginalia.wmsa.client.GsonFactory;
 import nu.marginalia.wmsa.edge.index.postings.SearchIndexControl;
 import nu.marginalia.wmsa.edge.index.query.IndexResultDomainDeduplicator;
@@ -101,7 +101,7 @@ public class EdgeIndexDomainQueryService {
 
     private OptionalInt lookUpWord(String s) {
         int ret = indexes.getLexiconReader().get(s);
-        if (ret == DictionaryHashMap.NO_VALUE) {
+        if (ret == OffHeapDictionaryHashMap.NO_VALUE) {
             return OptionalInt.empty();
         }
         return OptionalInt.of(ret);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/EdgeIndexLexiconService.java

@@ -3,7 +3,7 @@ package nu.marginalia.wmsa.edge.index.svc;
 import com.google.inject.Inject;
 import com.google.inject.Singleton;
 import com.google.protobuf.InvalidProtocolBufferException;
-import nu.marginalia.util.dict.DictionaryHashMap;
+import nu.marginalia.util.dict.OffHeapDictionaryHashMap;
 import nu.marginalia.wmsa.edge.converting.interpreter.instruction.DocumentKeywords;
 import nu.marginalia.wmsa.edge.converting.interpreter.instruction.KeywordListChunker;
 import nu.marginalia.wmsa.edge.index.IndexServicesFactory;
@@ -51,7 +51,7 @@ public class EdgeIndexLexiconService {
 
         final int wordId = lr.get(word);
 
-        if (DictionaryHashMap.NO_VALUE == wordId) {
+        if (OffHeapDictionaryHashMap.NO_VALUE == wordId) {
             response.status(404);
             return "";
         }
@@ -110,7 +110,7 @@ public class EdgeIndexLexiconService {
             String word = words[i];
 
             long id = keywordLexicon.getOrInsert(word);
-            if (id != DictionaryHashMap.NO_VALUE) {
+            if (id != OffHeapDictionaryHashMap.NO_VALUE) {
                 ids[putIdx++] = id;
                 ids[putIdx++] = meta[i];
             }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/EdgeIndexQueryService.java

@@ -12,7 +12,7 @@ import io.prometheus.client.Histogram;
 import it.unimi.dsi.fastutil.ints.IntArrayList;
 import it.unimi.dsi.fastutil.ints.IntList;
 import nu.marginalia.util.array.buffer.LongQueryBuffer;
-import nu.marginalia.util.dict.DictionaryHashMap;
+import nu.marginalia.util.dict.OffHeapDictionaryHashMap;
 import nu.marginalia.wmsa.client.GsonFactory;
 import nu.marginalia.wmsa.edge.index.postings.EdgeIndexQuerySearchTerms;
 import nu.marginalia.wmsa.edge.index.postings.IndexResultValuator;
@@ -115,11 +115,13 @@ public class EdgeIndexQueryService {
         TLongHashSet consideredUrlIds;
 
         public SearchQuery(EdgeSearchSpecification specsSet) {
-            this.fetchSize = specsSet.fetchSize;
-            this.budget = new IndexSearchBudget(specsSet.timeoutMs);
+            var limits = specsSet.queryLimits;
+
+            this.fetchSize = limits.fetchSize();
+            this.budget = new IndexSearchBudget(limits.timeoutMs());
             this.subqueries = specsSet.subqueries;
-            this.limitByDomain = specsSet.limitByDomain;
-            this.limitTotal = specsSet.limitTotal;
+            this.limitByDomain = limits.resultsByDomain();
+            this.limitTotal = limits.resultsTotal();
 
             this.consideredUrlIds = new TLongHashSet(fetchSize * 4);
 
@@ -127,6 +129,7 @@ public class EdgeIndexQueryService {
                     specsSet.quality,
                     specsSet.year,
                     specsSet.size,
+                    specsSet.rank,
                     getSearchSet(specsSet),
                     specsSet.queryStrategy);
         }
@@ -151,7 +154,7 @@ public class EdgeIndexQueryService {
                 }
             }
 
-            final var evaluator = new IndexResultValuator(indexes, results, subqueries);
+            final var evaluator = new IndexResultValuator(indexes, results, subqueries, queryParams);
 
             ArrayList<EdgeSearchResultItem> items = new ArrayList<>(results.size());
             ArrayList<EdgeSearchResultItem> refusedItems = new ArrayList<>(results.size());
@@ -293,7 +296,7 @@ public class EdgeIndexQueryService {
 
     private OptionalInt lookUpWord(String s) {
         int ret = indexes.getLexiconReader().get(s);
-        if (ret == DictionaryHashMap.NO_VALUE) {
+        if (ret == OffHeapDictionaryHashMap.NO_VALUE) {
             return OptionalInt.empty();
         }
         return OptionalInt.of(ret);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/EdgeIndexSearchSetsService.java

@@ -2,20 +2,20 @@ package nu.marginalia.wmsa.edge.index.svc;
 
 import com.google.inject.Inject;
 import com.google.inject.Singleton;
-import com.zaxxer.hikari.HikariDataSource;
-import gnu.trove.list.TIntList;
-import gnu.trove.list.array.TIntArrayList;
 import lombok.SneakyThrows;
-import nu.marginalia.util.ranking.BetterReversePageRank;
-import nu.marginalia.util.ranking.BetterStandardPageRank;
-import nu.marginalia.util.ranking.RankingDomainFetcher;
+import nu.marginalia.wmsa.edge.index.ranking.ReversePageRank;
+import nu.marginalia.wmsa.edge.index.ranking.StandardPageRank;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultHashMapAccumulator;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcher;
+import nu.marginalia.wmsa.edge.index.ranking.accumulator.RankingResultBitSetAccumulator;
 import nu.marginalia.wmsa.edge.index.IndexServicesFactory;
 import nu.marginalia.wmsa.edge.index.model.RankingSettings;
+import nu.marginalia.wmsa.edge.index.postings.DomainRankings;
+import nu.marginalia.wmsa.edge.index.ranking.data.RankingDomainFetcherForSimilarityData;
 import nu.marginalia.wmsa.edge.index.svc.searchset.RankingSearchSet;
 import nu.marginalia.wmsa.edge.index.svc.searchset.SearchSet;
 import nu.marginalia.wmsa.edge.index.svc.searchset.SearchSetAny;
 import nu.marginalia.wmsa.edge.index.svc.searchset.SearchSetIdentifier;
-import org.roaringbitmap.RoaringBitmap;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -23,137 +23,47 @@ import java.io.IOException;
 
 @Singleton
 public class EdgeIndexSearchSetsService {
-    private final HikariDataSource dataSource;
-    private RankingDomainFetcher rankingDomains;
-    private final RankingSettings rankingSettings;
     private final Logger logger = LoggerFactory.getLogger(getClass());
+    private final RankingDomainFetcher rankingDomains;
+    private final RankingDomainFetcher similarityDomains;
+    private final RankingSettings rankingSettings;
 
-    private final SearchSet anySet = new SearchSetAny();
+
+    // Below are binary indices that are used to constrain a search
     private volatile RankingSearchSet retroSet;
     private volatile RankingSearchSet smallWebSet;
     private volatile RankingSearchSet academiaSet;
+    private final SearchSet anySet = new SearchSetAny();
+
+    // The ranking value of the domains used in sorting the domains
+    private volatile DomainRankings domainRankings = new DomainRankings();
 
     @Inject
-    public EdgeIndexSearchSetsService(HikariDataSource dataSource,
-                                      RankingDomainFetcher rankingDomains,
+    public EdgeIndexSearchSetsService(RankingDomainFetcher rankingDomains,
+                                      RankingDomainFetcherForSimilarityData similarityDomains,
                                       RankingSettings rankingSettings,
                                       IndexServicesFactory servicesFactory) throws IOException {
-        this.dataSource = dataSource;
+
         this.rankingDomains = rankingDomains;
+
+        if (similarityDomains.hasData()) {
+            this.similarityDomains = similarityDomains;
+        }
+        else {
+            // on test environments the cosine similarity graph may not be present
+            logger.info("Domain similarity is not present, falling back on link graph");
+            this.similarityDomains = rankingDomains;
+        }
+
         this.rankingSettings = rankingSettings;
 
         smallWebSet = new RankingSearchSet(SearchSetIdentifier.SMALLWEB, servicesFactory.getSearchSetsBase().resolve("small-web.dat"));
         academiaSet = new RankingSearchSet(SearchSetIdentifier.ACADEMIA, servicesFactory.getSearchSetsBase().resolve("academia.dat"));
         retroSet = new RankingSearchSet(SearchSetIdentifier.RETRO, servicesFactory.getSearchSetsBase().resolve("retro.dat"));
-
-        logger.info("SearchIndexDao ranking settings = {}", rankingSettings);
     }
 
-    public void recalculateAll() {
-        updateAcademiaDomains();
-        updateRetroDomains();
-        updateSmallWebDomains();
-    }
-
-    @SneakyThrows
-    public RoaringBitmap goodUrls() {
-        RoaringBitmap domains = new RoaringBitmap();
-        RoaringBitmap urls = new RoaringBitmap();
-
-        try (var connection = dataSource.getConnection()) {
-            try (var stmt = connection.prepareStatement("SELECT ID FROM EC_DOMAIN WHERE DOMAIN_ALIAS IS NULL AND IS_ALIVE")) {
-                stmt.setFetchSize(10_000);
-                var rsp = stmt.executeQuery();
-                while (rsp.next()) {
-                    domains.add(rsp.getInt(1));
-                }
-            }
-
-            // For some reason, doing this "INNER JOIN" in Java is significantly faster than doing it in SQL
-            try (var stmt = connection.prepareStatement("SELECT ID,DOMAIN_ID FROM EC_URL WHERE VISITED AND EC_URL.STATE='OK'")) {
-                stmt.setFetchSize(10_000);
-                var rsp = stmt.executeQuery();
-                while (rsp.next()) {
-                    if (domains.contains(rsp.getInt(2))) {
-                        urls.add(rsp.getInt(1));
-                    }
-                }
-            }
-
-        }
-
-        return urls;
-    }
-
-    @SneakyThrows
-    public void updateRetroDomains() {
-        var spr = new BetterStandardPageRank(rankingDomains,rankingSettings.retro.toArray(String[]::new));
-        var data = spr.pageRankWithPeripheralNodes(spr.size() / 2);
-
-        synchronized (this) {
-            retroSet = new RankingSearchSet(SearchSetIdentifier.RETRO, retroSet.source, data);
-            retroSet.write();
-        }
-    }
-
-    @SneakyThrows
-    public void updateSmallWebDomains() {
-        var rpr = new BetterReversePageRank(rankingDomains,  rankingSettings.small.toArray(String[]::new));
-        rpr.setMaxKnownUrls(750);
-        var data = rpr.pageRankWithPeripheralNodes(rpr.size());
-
-        synchronized (this) {
-            smallWebSet = new RankingSearchSet(SearchSetIdentifier.SMALLWEB, smallWebSet.source, data);
-            smallWebSet.write();
-        }
-    }
-
-    @SneakyThrows
-    public void updateAcademiaDomains() {
-        var spr =  new BetterStandardPageRank(rankingDomains,  rankingSettings.academia.toArray(String[]::new));
-        var data = spr.pageRankWithPeripheralNodes(spr.size()/2);
-
-        synchronized (this) {
-            academiaSet = new RankingSearchSet(SearchSetIdentifier.ACADEMIA, academiaSet.source, data);
-            academiaSet.write();
-        }
-    }
-
-    @SneakyThrows
-    public TIntList getStandardDomains() {
-        TIntArrayList results = new TIntArrayList();
-
-        try (var connection = dataSource.getConnection();
-             var stmt = connection.prepareStatement(
-            """
-            SELECT ID FROM EC_DOMAIN 
-            WHERE INDEXED>0 
-            AND STATE='ACTIVE' 
-            AND DOMAIN_ALIAS IS NULL 
-            ORDER BY ID ASC
-            """);
-        ) {
-            var rs = stmt.executeQuery();
-            while (rs.next()) {
-                results.add(rs.getInt(1));
-            }
-        }
-        return results;
-
-    }
-
-    @SneakyThrows
-    public TIntList getSpecialDomains() {
-        TIntArrayList results = new TIntArrayList();
-        try (var connection = dataSource.getConnection();
-             var stmt = connection.prepareStatement("SELECT ID FROM EC_DOMAIN WHERE STATE='SPECIAL'")
-        ) {
-            var rs = stmt.executeQuery();
-            while (rs.next()) {
-                results.add(rs.getInt(1));
-            }
-        }
-        return results;
+    public DomainRankings getDomainRankings() {
+        return domainRankings;
     }
 
     public SearchSet getSearchSetByName(SearchSetIdentifier searchSetIdentifier) {
@@ -167,4 +77,54 @@ public class EdgeIndexSearchSetsService {
             case SMALLWEB -> smallWebSet;
         };
     }
+
+    public void recalculateAll() {
+        updateAcademiaDomainsSet();
+        updateRetroDomainsSet();
+        updateSmallWebDomainsSet();
+        updateDomainRankings();
+    }
+
+    private void updateDomainRankings() {
+        var spr = new StandardPageRank(similarityDomains, rankingSettings.retro.toArray(String[]::new));
+
+        var ranks = spr.pageRankWithPeripheralNodes(spr.size() / 2, () -> new RankingResultHashMapAccumulator(100_000));
+        synchronized (this) {
+            domainRankings = new DomainRankings(ranks);
+        }
+    }
+
+    @SneakyThrows
+    public void updateRetroDomainsSet() {
+        var spr = new StandardPageRank(similarityDomains, rankingSettings.retro.toArray(String[]::new));
+        var data = spr.pageRankWithPeripheralNodes(spr.size() / 2, RankingResultBitSetAccumulator::new);
+
+        synchronized (this) {
+            retroSet = new RankingSearchSet(SearchSetIdentifier.RETRO, retroSet.source, data);
+            retroSet.write();
+        }
+    }
+
+    @SneakyThrows
+    public void updateSmallWebDomainsSet() {
+        var rpr = new ReversePageRank(similarityDomains,  rankingSettings.small.toArray(String[]::new));
+        rpr.setMaxKnownUrls(750);
+        var data = rpr.pageRankWithPeripheralNodes(rpr.size(), RankingResultBitSetAccumulator::new);
+
+        synchronized (this) {
+            smallWebSet = new RankingSearchSet(SearchSetIdentifier.SMALLWEB, smallWebSet.source, data);
+            smallWebSet.write();
+        }
+    }
+
+    @SneakyThrows
+    public void updateAcademiaDomainsSet() {
+        var spr =  new StandardPageRank(similarityDomains,  rankingSettings.academia.toArray(String[]::new));
+        var data = spr.pageRankWithPeripheralNodes(spr.size()/2, RankingResultBitSetAccumulator::new);
+
+        synchronized (this) {
+            academiaSet = new RankingSearchSet(SearchSetIdentifier.ACADEMIA, academiaSet.source, data);
+            academiaSet.write();
+        }
+    }
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/searchset/RankingSearchSet.java

@@ -9,21 +9,37 @@ import java.nio.file.Files;
 import java.nio.file.Path;
 import java.nio.file.StandardOpenOption;
 
+/** A serializable bit map of domains
+ *
+ * @see SearchSetIdentifier
+ *
+ * */
 public class RankingSearchSet implements SearchSet {
 
     private final RoaringBitmap set;
     public final SearchSetIdentifier identifier;
     public final Path source;
 
+    public RankingSearchSet(SearchSetIdentifier identifier, Path source, RoaringBitmap set) {
+        this.identifier = identifier;
+        this.source = source;
+        this.set = set;
+    }
+
     public RankingSearchSet(SearchSetIdentifier identifier, Path source) throws IOException {
         this.identifier = identifier;
         this.source = source;
-        set = new RoaringBitmap();
 
         if (!Files.exists(source)) {
-            return;
+            set = new RoaringBitmap();
         }
+        else {
+            set = load(source);
+        }
+    }
 
+    private static RoaringBitmap load(Path source) throws IOException {
+        var set = new RoaringBitmap();
         try (var ds = new DataInputStream(Files.newInputStream(source, StandardOpenOption.READ))) {
             for (;;) {
                 try {
@@ -32,12 +48,7 @@ public class RankingSearchSet implements SearchSet {
                 catch (IOException ex) { break; }
             }
         }
-    }
-
-    public RankingSearchSet(SearchSetIdentifier identifier, Path source, RoaringBitmap set) {
-        this.identifier = identifier;
-        this.source = source;
-        this.set = set;
+        return set;
     }
 
     @Override
@@ -46,7 +57,11 @@ public class RankingSearchSet implements SearchSet {
     }
 
     public void write() throws IOException {
-        try (var ds = new DataOutputStream(Files.newOutputStream(source, StandardOpenOption.WRITE,  StandardOpenOption.CREATE, StandardOpenOption.TRUNCATE_EXISTING))) {
+        try (var ds = new DataOutputStream(Files.newOutputStream(source,
+                StandardOpenOption.WRITE,
+                StandardOpenOption.CREATE,
+                StandardOpenOption.TRUNCATE_EXISTING)))
+        {
             for (var iter = set.getIntIterator(); iter.hasNext();) {
                 ds.writeInt(iter.next());
             }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/searchset/SearchSetIdentifier.java

@@ -1,5 +1,12 @@
 package nu.marginalia.wmsa.edge.index.svc.searchset;
 
+import nu.marginalia.wmsa.edge.search.model.EdgeSearchProfile;
+
+/** Identifies a RankingSearchSet, associated with an EdgeSearchProfile
+ *
+ * @see RankingSearchSet
+ * @see EdgeSearchProfile
+ * */
 public enum SearchSetIdentifier {
     NONE,
     RETRO,

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/svc/searchset/SmallSearchSet.java

@@ -13,8 +13,8 @@ public class SmallSearchSet implements SearchSet {
     }
 
     @Override
-    public boolean contains(int urlId) {
-        return entries.contains(urlId);
+    public boolean contains(int domainId) {
+        return entries.contains(domainId);
     }
 
     public String toString() {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/integration/stackoverflow/StackOverflowPostProcessor.java

@@ -2,7 +2,7 @@ package nu.marginalia.wmsa.edge.integration.stackoverflow;
 
 import com.google.inject.Inject;
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.KeywordMetadata;
 import nu.marginalia.wmsa.edge.converting.processor.logic.LinkParser;
 import nu.marginalia.wmsa.edge.integration.model.BasicDocumentData;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/integration/wikipedia/WikipediaProcessor.java

@@ -1,7 +1,7 @@
 package nu.marginalia.wmsa.edge.integration.wikipedia;
 
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.KeywordMetadata;
 import nu.marginalia.wmsa.edge.converting.processor.logic.LinkParser;
 import nu.marginalia.wmsa.edge.integration.model.BasicDocumentData;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/model/EdgeDomain.java

@@ -11,8 +11,6 @@ import java.util.regex.Pattern;
 @Getter @Setter @Builder
 public class EdgeDomain {
 
-    private static final Predicate<String> ipPatternTest = Pattern.compile("[\\d]{1,3}\\.[\\d]{1,3}\\.[\\d]{1,3}\\.[\\d]{1,3}").asMatchPredicate();
-    private static final Predicate<String> govListTest = Pattern.compile(".*\\.(ac|co|org|gov|edu|com)\\.[a-z]{2}").asMatchPredicate();
 
     @Nonnull
     public final String subDomain;
@@ -27,7 +25,7 @@ public class EdgeDomain {
 
         var dot = host.lastIndexOf('.');
 
-        if (dot < 0 || ipPatternTest.test(host)) { // IPV6 >.>
+        if (dot < 0 || looksLikeAnIp(host)) { // IPV6 >.>
             subDomain = "";
             domain = host;
         }
@@ -38,7 +36,7 @@ public class EdgeDomain {
                 domain = host;
             }
             else {
-                if (govListTest.test(host))
+                if (looksLikeGovTld(host))
                 { // Capture .ac.jp, .co.uk
                     int dot3 = host.substring(0, dot2).lastIndexOf('.');
                     if (dot3 >= 0) {
@@ -59,6 +57,35 @@ public class EdgeDomain {
         }
     }
 
+    private static final Predicate<String> govListTest = Pattern.compile(".*\\.(ac|co|org|gov|edu|com)\\.[a-z]{2}").asMatchPredicate();
+    private boolean looksLikeGovTld(String host) {
+        if (host.length() < 8)
+            return false;
+        int cnt = 0;
+        for (int i = host.length() - 7; i < host.length(); i++) {
+            if (host.charAt(i) == '.')
+                cnt++;
+        }
+        return cnt >= 2 && govListTest.test(host);
+    }
+
+
+    private static final Predicate<String> ipPatternTest = Pattern.compile("[\\d]{1,3}\\.[\\d]{1,3}\\.[\\d]{1,3}\\.[\\d]{1,3}").asMatchPredicate();
+
+    private boolean looksLikeAnIp(String host) {
+        if (host.length() < 7)
+            return false;
+
+        char firstChar = host.charAt(0);
+        int lastChar = host.charAt(host.length() - 1);
+
+        return Character.isDigit(firstChar)
+                && Character.isDigit(lastChar)
+                && ipPatternTest.test(host);
+    }
+
+
+
     public EdgeUrl toRootUrl() {
         // Set default protocol to http, as most https websites redirect http->https, but few http websites redirect https->http
         return new EdgeUrl("http", this, null, "/", null);

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/model/search/EdgeSearchResultKeywordScore.java

@@ -24,6 +24,11 @@ public record EdgeSearchResultKeywordScore(int set,
             sum +=  20;
         }
 
+        int rank = EdgePageDocumentsMetadata.decodeRank(encodedDocMetadata) - 13;
+        if (rank < 0)
+            sum += rank / 2;
+        else
+            sum += rank / 4;
 
         return sum;
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/model/search/EdgeSearchSpecification.java

@@ -1,6 +1,7 @@
 package nu.marginalia.wmsa.edge.model.search;
 
 import lombok.*;
+import nu.marginalia.wmsa.edge.index.model.QueryLimits;
 import nu.marginalia.wmsa.edge.index.model.QueryStrategy;
 import nu.marginalia.wmsa.edge.index.svc.searchset.SearchSetIdentifier;
 import nu.marginalia.wmsa.edge.model.search.domain.SpecificationLimit;
@@ -9,23 +10,18 @@ import java.util.List;
 
 @ToString @Getter @Builder @With @AllArgsConstructor
 public class EdgeSearchSpecification {
-
     public List<EdgeSearchSubquery> subqueries;
     public List<Integer> domains;
     public SearchSetIdentifier searchSetIdentifier;
 
-    public final int limitByDomain;
-    public final int limitTotal;
-
     public final String humanQuery;
 
-    public final int timeoutMs;
-    public final int fetchSize;
-
     public final SpecificationLimit quality;
     public final SpecificationLimit year;
     public final SpecificationLimit size;
+    public final SpecificationLimit rank;
 
+    public final QueryLimits queryLimits;
     public final QueryStrategy queryStrategy;
 
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/QueryFactory.java

@@ -6,6 +6,7 @@ import nu.marginalia.util.language.WordPatterns;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
 import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
+import nu.marginalia.wmsa.edge.index.model.QueryLimits;
 import nu.marginalia.wmsa.edge.index.model.QueryStrategy;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchSpecification;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchSubquery;
@@ -84,6 +85,8 @@ public class QueryFactory {
         List<String> problems = new ArrayList<>();
         String domain = null;
 
+        QueryStrategy queryStrategy = QueryStrategy.AUTO;
+
         var basicQuery = queryParser.parse(query);
 
         if (basicQuery.size() >= 8) {
@@ -94,6 +97,7 @@ public class QueryFactory {
         SpecificationLimit qualityLimit = profile.getQualityLimit();
         SpecificationLimit year = profile.getYearLimit();
         SpecificationLimit size = profile.getSizeLimit();
+        SpecificationLimit rank = SpecificationLimit.none();
 
         for (Token t : basicQuery) {
             if (t.type == TokenType.QUOT_TERM || t.type == TokenType.LITERAL_TERM) {
@@ -113,6 +117,12 @@ public class QueryFactory {
             if (t.type == TokenType.SIZE_TERM) {
                 size = parseSpecificationLimit(t.str);
             }
+            if (t.type == TokenType.RANK_TERM) {
+                rank = parseSpecificationLimit(t.str);
+            }
+            if (t.type == TokenType.QS_TERM) {
+                queryStrategy = parseQueryStrategy(t.str);
+            }
         }
 
         var queryPermutations = queryParser.permuteQueriesNew(basicQuery);
@@ -148,6 +158,8 @@ public class QueryFactory {
                     case QUALITY_TERM:
                     case YEAR_TERM:
                     case SIZE_TERM:
+                    case RANK_TERM:
+                    case QS_TERM:
                         break; //
                     case NEAR_TERM:
                         near = t.str;
@@ -179,25 +191,25 @@ public class QueryFactory {
             }
         }
 
+        int domainLimit;
+        if (domain != null) {
+            domainLimit = 100;
+        } else {
+            domainLimit = 2;
+        }
+
         EdgeSearchSpecification.EdgeSearchSpecificationBuilder specsBuilder = EdgeSearchSpecification.builder()
                 .subqueries(subqueries)
-                .limitTotal(100)
+                .queryLimits(new QueryLimits(domainLimit, 100, 250, 4096))
                 .humanQuery(query)
-                .timeoutMs(250)
-                .fetchSize(4096)
                 .quality(qualityLimit)
                 .year(year)
                 .size(size)
+                .rank(rank)
                 .domains(domains)
-                .queryStrategy(QueryStrategy.AUTO)
+                .queryStrategy(queryStrategy)
                 .searchSetIdentifier(profile.searchSetIdentifier);
 
-        if (domain != null) {
-            specsBuilder = specsBuilder.limitByDomain(100);
-        } else {
-            specsBuilder = specsBuilder.limitByDomain(2);
-        }
-
         EdgeSearchSpecification specs = specsBuilder.build();
 
         return new EdgeSearchQuery(specs, searchTermsHuman, domain);
@@ -210,10 +222,10 @@ public class QueryFactory {
         if (startChar == '=') {
             return SpecificationLimit.equals(val);
         }
-        else if (startChar == '<'){
+        else if (startChar == '<') {
             return SpecificationLimit.lessThan(val);
         }
-        else if (startChar == '>'){
+        else if (startChar == '>') {
             return SpecificationLimit.greaterThan(val);
         }
         else {
@@ -221,6 +233,17 @@ public class QueryFactory {
         }
     }
 
+    private QueryStrategy parseQueryStrategy(String str) {
+        return switch (str.toUpperCase()) {
+            case "RF_TITLE" -> QueryStrategy.REQUIRE_FIELD_TITLE;
+            case "RF_SUBJECT" -> QueryStrategy.REQUIRE_FIELD_SUBJECT;
+            case "RF_SITE" -> QueryStrategy.REQUIRE_FIELD_SITE;
+            case "SENTENCE" -> QueryStrategy.SENTENCE;
+            case "TOPIC" -> QueryStrategy.TOPIC;
+            default -> QueryStrategy.AUTO;
+        };
+    }
+
     private String normalizeDomainName(String str) {
         return str.toLowerCase();
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/QueryParser.java

@@ -93,6 +93,10 @@ public class QueryParser {
                 entity.replace(new Token(TokenType.YEAR_TERM, t.str.substring(4), t.displayStr));
             } else if (t.str.startsWith("size") && t.str.matches("size[=><]\\d+")) {
                 entity.replace(new Token(TokenType.SIZE_TERM, t.str.substring(4), t.displayStr));
+            } else if (t.str.startsWith("rank") && t.str.matches("rank[=><]\\d+")) {
+                entity.replace(new Token(TokenType.RANK_TERM, t.str.substring(4), t.displayStr));
+            } else if (t.str.startsWith("qs=")) {
+                entity.replace(new Token(TokenType.QS_TERM, t.str.substring(3), t.displayStr));
             } else if (t.str.contains(":")) {
                 entity.replace(new Token(TokenType.ADVICE_TERM, t.str, t.displayStr));
             }
@@ -506,8 +510,11 @@ enum TokenType implements Predicate<Token> {
     QUALITY_TERM,
     YEAR_TERM,
     SIZE_TERM,
+    RANK_TERM,
     NEAR_TERM,
 
+    QS_TERM,
+
     QUOT,
     MINUS,
     QMARK,

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/QueryVariants.java

@@ -8,7 +8,7 @@ import lombok.Getter;
 import lombok.ToString;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.KeywordExtractor;
-import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.util.language.processing.sentence.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordSpan;
 import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
@@ -25,12 +25,12 @@ public class QueryVariants {
 
     private final Logger logger = LoggerFactory.getLogger(getClass());
     private final KeywordExtractor keywordExtractor;
-    private final SentenceExtractor sentenceExtractor;
     private final TermFrequencyDict dict;
     private final PorterStemmer ps = new PorterStemmer();
 
     private final NGramBloomFilter nGramBloomFilter;
     private final EnglishDictionary englishDictionary;
+    private final ThreadLocal<SentenceExtractor> sentenceExtractor;
 
     @Inject
     public QueryVariants(LanguageModels lm,
@@ -40,7 +40,7 @@ public class QueryVariants {
         this.nGramBloomFilter = nGramBloomFilter;
         this.englishDictionary = englishDictionary;
         this.keywordExtractor = new KeywordExtractor();
-        this.sentenceExtractor = new SentenceExtractor(lm);
+        this.sentenceExtractor = ThreadLocal.withInitial(() -> new SentenceExtractor(lm));
         this.dict = dict;
     }
 
@@ -78,10 +78,8 @@ public class QueryVariants {
 
         final TreeMap<Integer, List<WordSpan>> byStart = new TreeMap<>();
 
-        logger.debug("Q: {}", query);
-        logger.debug("QAS: {}", joinedQuery);
-
-        var sentence = sentenceExtractor.extractSentence(joinedQuery.joinedQuery);
+        var se = sentenceExtractor.get();
+        var sentence = se.extractSentence(joinedQuery.joinedQuery);
 
         for (int i = 0; i < sentence.posTags.length; i++) {
             if (sentence.posTags[i].startsWith("N") || sentence.posTags[i].startsWith("V")) {