Skip to content

perf(vector/search): PQ rerank pipeline not exposed: PQ → SQ → f32 not possible #673

Description

@mosuka

Round-3 perf push sub-issue (tracked under umbrella #536).

[M] PQ rerank pipeline not exposed: PQ → SQ → f32 not possible

  • Where: HNSW Stage-2 rerank (hnsw/searcher.rs:493-522) widens int8 → f32 only. No
    path lets PQ candidates be reranked first against int8 then f32.
  • Current behavior: Either PQ-only (Stage 3) or SQ-only (Stage 1) + optional f32 rerank
    (Stage 2). No 3-stage chain.
  • Reference precedent: LanceDB IVF_PQ → row-group rerank; FAISS IndexRefineFlat over
    IndexPQ; ScaNN multi-stage.
  • Suggested direction: Generic RerankPipeline (see L-sized rerank item) lets a
    configurable stack [PqScorer, SqScorer, F32Scorer]. Each stage narrows by factor,
    emits to next.
  • Risk / scope: Large; part of the rerank refactor.

ID: VS-30 — see ~/.claude/tasks/laurus/20260523_perf_round3_audit/task_list.md for the full Round-3 issue list.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions