LSM_SPARSE_VECTOR: expose weightQuantization in index METADATA (format already supports FP32/FP16/INT8)

[tae898]

Summary

LSM_SPARSE_VECTOR hard-wires INT8 posting-weight quantization: SegmentParameters.Builder defaults to WeightQuantization.INT8, and no CREATE INDEX ... METADATA key exposes the choice — even though SegmentFormat.WeightQuantization already defines FP32, FP16, and INT8.

The dense index already has exactly this knob (LSMVectorIndexMetadata.quantizationType via METADATA {"quantization": ...}, default NONE), so this would bring the sparse index to parity with its dense sibling.

Why it matters

INT8 is a good default — in our controlled benchmarks (10M docs, 30k dims, SPLADE-shaped, exact brute-force ground truth) it holds recall@10 at 0.9930 with perfect determinism across 5 repeated runs. But:

• users who need exact scoring (like Qdrant/Milvus deliver with f32 postings, recall 1.0 in the same benchmark) have no way to opt into FP32;
• there is no way to quantify the INT8 tradeoff (recall vs latency vs disk) on one's own workload, because both sides of the comparison can't be built.

Suggested shape

CREATE INDEX ON Doc (tokens, weights) LSM_SPARSE_VECTOR
METADATA { "dimensions": 30000, "weightQuantization": "FP32" }

Default stays INT8. Happy to contribute benchmark numbers for the INT8-vs-FP32 comparison once the knob exists.

[lvca]

Implemented and merged into main (commit 8a9f683).

You can now pick the posting-weight quantization when creating a LSM_SPARSE_VECTOR index, exactly as suggested:

CREATE INDEX ON Doc (tokens, weights) LSM_SPARSE_VECTOR
METADATA { "dimensions": 30000, "weightQuantization": "FP32" }

Accepted values are INT8 (default, most compact), FP16, and FP32 (exact scoring). The setting is saved with the index and controls how new data is written; indexes you already have keep working unchanged. There is also an equivalent Java API method, withWeightQuantization(...).

Ships in 26.7.2. Benchmark numbers for the INT8-vs-FP32 comparison would be very welcome once you get a chance to try it.

[tae898]

Confirmed in our benchmarks: the weightQuantization knob (INT8 default vs FP32) is exposed via index METADATA and both quantizations run cleanly at the 10M-doc tier (INT8 recall@10 ~0.993, FP32 1.0). Thanks.