Add fallback support to Lucene104ScalarQuantizedVectorsFormat getFloatVectorValues when there are no full-precision vectors present #15415
Conversation
…tVectorValues when there are no full-precision vectors present
benwtrent
left a comment
benwtrent
left a comment
There was a problem hiding this comment.
Looking good! I am sorry we missed this with the new format. Thank you for taking care of it.
lucene/CHANGES.txt
Outdated
| (Ben Trent) | ||
|
|
||
| * GITHUB#15415: Add fallback support to Lucene104ScalarQuantizedVectorsFormat getFloatVectorValues when there are | ||
| no full-precision vectors present |
There was a problem hiding this comment.
Please add your name for posterity :)
There was a problem hiding this comment.
Oh I missed it. Fixed in next revision.
| } | ||
|
|
||
| OffHeapScalarQuantizedFloatVectorValues( | ||
| boolean isQuerySide, |
There was a problem hiding this comment.
We should never allow this "querySide" thing. Even as it is now, it wouldn't work.
| // unpack bytes | ||
| switch (encoding) { | ||
| case PACKED_NIBBLE -> | ||
| OffHeapScalarQuantizedVectorValues.unpackNibbles(byteValue, unpackedByteVectorValue); | ||
| case SINGLE_BIT_QUERY_NIBBLE -> | ||
| OptimizedScalarQuantizer.unpackBinary(byteValue, unpackedByteVectorValue); | ||
| case UNSIGNED_BYTE, SEVEN_BIT -> { | ||
| deQuantize(byteValue, vectorValue, encoding.getBits(), correctiveValues, centroid); | ||
| lastOrd = targetOrd; | ||
| return vectorValue; | ||
| } | ||
| } |
There was a problem hiding this comment.
If this was "query side" it wouldn't work. Consequently, I think this query side thing should go away.
I think this piece is great if we always assume document quantized bits.
| byte[] quantized, | ||
| float[] dequantized, | ||
| byte bits, | ||
| float[] correctiveValues, |
There was a problem hiding this comment.
maybe make this specifically lower upper thing instead of this array?
| byte[] quantized, | ||
| float[] dequantized, | ||
| byte bits, | ||
| float[] correctiveValues, |
There was a problem hiding this comment.
lets name these lowerInterval and upperInterval
There was a problem hiding this comment.
Renamed in next revision.
|
Thanks @benwtrent for such quick review!! |
2 participants
Description
Add fallback support to Lucene104ScalarQuantizedVectorsFormat.getFloatVectorValues() when there are no full-precision vectors present. As part of this PR, we added this support in Lucene99ScalarQuantizedVectorsFormat but it got missed in new vector codec. This PR is trying to add back that support.