pg_retrieval_engine | Hybrid Retrieval for PostgreSQL

项目内容 Project Scope

FAISS in PostgreSQL

在 PostgreSQL backend 内运行 FAISS ANN 索引，支持 HNSW、IVFFlat、IVFPQ。

Runs FAISS ANN indexes inside a PostgreSQL backend with HNSW, IVFFlat, and IVFPQ.

RRF SQL

融合 pgvector 向量排名与 PostgreSQL tsvector 全文排名。

Fuses pgvector rankings with PostgreSQL tsvector full-text rankings.

Evaluation

离线计算 Recall@K、NDCG@K、P95/P99 latency，用于验证融合效果。

Computes Recall@K, NDCG@K, and P95/P99 latency to validate fusion quality.

Observability

暴露查询调用量、耗时、候选深度、batch 参数和自动调参状态。

Exposes call counts, timings, candidate depth, batch knobs, and autotune state.

快速使用 Quick Usage

CREATE EXTENSION vector;
CREATE EXTENSION pg_retrieval_engine;

SELECT pg_retrieval_engine_index_create(
  'docs_hnsw', 768, 'cosine', 'hnsw',
  '{"m":32,"ef_construction":200,"ef_search":64}'::jsonb,
  'cpu'
);

SELECT *
FROM pg_retrieval_engine_hybrid_search(
  'documents'::regclass,
  'id',
  'embedding',
  'search_vector',
  '[0.1,0.2,0.3,0.4]'::vector,
  plainto_tsquery('simple', 'vector database'),
  20,
  '{"vector_k":100,"fts_k":100,"rrf_k":60}'::jsonb
);

量化指标 Quantitative Metrics

Module	Direction	Required metrics	Current target/result
FAISS	ANN	Recall@10, avg latency, P95/P99, speedup	CPU target >= 5x; local HNSW 11.32x, IVFFlat 10.31x
RRF SQL	Hybrid fusion	Recall@K, NDCG@K, P95/P99	Report vector-only, FTS-only, and RRF together
Autotune	Runtime tuning	before/after Recall@K, NDCG@K, P95/P99	Validate with fixed qrels before accepting changes