On main, a 9.x/10.x int7/int4 scalar-quantized segment whose vectors are all identical (stored quantiles equal) is decoded with the min == max widening that #14049 added, not the quantizer that wrote it, so its vectors read 127 (int7) or 15 (int4) units lower in every dimension and exact matches score about 0 and rank last (unreleased, 11.0 only)

Finding LFV-841Area: codecs / legacy scalar-quantized vectors (util/quantization ScalarQuantizer, backward-codecs Lucene99ScalarQuantizedVectorsReader)Priority: lowImpact: silent wrong resultsNovelty: new

Finding page for LFV-841

AI-assisted finding, confirmed by automated checks, not yet triaged by a human Lucene committer. This site files nothing upstream.

What it shows

You index a field with the Lucene99 HNSW scalar-quantized format (int7, EUCLIDEAN) on 10.3.2 or 9.12.3, one segment holds only documents whose vector is [1,1,1,1], so the writer stores the quantiles (1, 1) and every code 0; you then open the unchanged index on main and search for [1,1,1,1]. Every release reader tested (9.9.2 to 10.6.0) decodes code 0 as 1.0, as the ScalarQuantizer formula for the stored quantiles says, and ranks those exact matches first with score 1.0. On main the reader rebuilds the quantizer through the min == max widening added by #14049, decodes those vectors as [-126,-126,-126,-126], and the exact matches score 1.5e-5 and rank last behind a document with half the similarity; int4 fields decode 15 units low, and DOT_PRODUCT and COSINE exact matches score 0.0.

Expected vs actual

Expected: a0, a1, a2 (the exact matches) rank first with score 1.0; b0 follows with 0.492.
Actual: b0 ranks first (0.492); a0, a1, a2 score 0.0000153 and rank last.

Files

src/org/apache/lucene/backward_index/TestOldUniformQuantizedSegment.java (JUnit test)

Run it

Tested on Lucene main e1faeba08bde (2026-11-02): still reproduces.
Tested on Lucene main a9779ef3c193 (2026-10-26): reproduced.
The tests embed their indexes as base64 zips, so nothing has to be generated to run them: the src/ test holds an index written through the public API of the 10.3.2 release jars and the same index written by 9.12.3, and the test shown on the finding page holds the 10.3.2 one. The generator that wrote them (GenUniformSqIndex, named in the src/ test's comments) is not included: it needs those release jars and does not compile against Lucene main, which can no longer write this format.
How to read the outcome: a failing test is the finding reproducing (the failure message states expected against actual).
Seeds: a rerun with the same seed replays the same random choices. Assertions off (-da -dsa -Dtests.asserts=false) is the production-like setting.

# Lucene checkout at the tested commit, built with: ./gradlew --no-daemon assemble   (JDK 25)
LUCENE=/path/to/lucene; V=12.0.0-SNAPSHOT
CP=$LUCENE/lucene/core/build/libs/lucene-core-$V.jar
CP=$CP:$LUCENE/lucene/codecs/build/libs/lucene-codecs-$V.jar
CP=$CP:$LUCENE/lucene/test-framework/build/libs/lucene-test-framework-$V.jar
CP=$CP:$LUCENE/lucene/backward-codecs/build/libs/lucene-backward-codecs-$V.jar
CP=$CP:randomizedtesting-runner-2.9.2.jar:junit-4.13.2.jar:hamcrest-3.0.jar   # test-framework dependencies (Maven Central)
CP=$CP:junit-platform-console-standalone-6.1.3.jar   # JUnit Platform console launcher (Maven Central)

javac -d classes -cp "$CP" $(find src -name '*.java')

# run 1: assertions on, seed DEADBEEF
java -ea -esa --add-modules jdk.incubator.vector --enable-native-access=ALL-UNNAMED -Dtests.seed=DEADBEEF \
     -cp "classes:$CP" org.junit.platform.console.ConsoleLauncher \
     execute --select-class org.apache.lucene.backward_index.TestOldUniformQuantizedSegment

Gradle alternative: copy the test into the matching module's src/test/ tree of the checkout and run ./gradlew -p lucene/<module> test --tests <class> -Ptests.seed=<seed> (add -Ptests.asserts=false for a run without assertions).

Files shown on the finding page

TestOldUniformQuantizedSegment.java

1 other file(s) of the recorded repro are not exported.

Formal models and scripts this finding's part of the hunt used: legacy-int7-int4-scalar-quantized-scoring-fp-smt

Feedback

Open an issue in lucene-ai-bug-hunt. Please do not open issues or pull requests in apache/lucene about this finding unless a Lucene committer agrees it should be filed.

Snapshot of 2026-11-04. AI-assisted; not yet triaged by a human Lucene committer. Feedback: open an issue in lucene-ai-bug-hunt.