pgvector-search（pg 向量搜索）

PGVector Hybrid Search

Production-grade semantic + keyword search using PostgreSQL

Overview

Architecture:

Query
  |
[Generate embedding] --> Vector Search (PGVector) --> Top 30 results
  |
[Generate ts_query]  --> Keyword Search (BM25)    --> Top 30 results
  |
[Reciprocal Rank Fusion (RRF)] --> Merge & re-rank --> Top 10 final results

When to use this skill:

• Building semantic search (RAG, knowledge bases, recommendations)
• Implementing hybrid retrieval (vector + keyword)
• Optimizing PGVector performance
• Working with large document collections (1M+ chunks)

Quick Reference

Search Type Comparison

Index Comparison

Recommendation: Use HNSW for production (scales to millions).

RRF Formula

rrf_score = 1/(k + vector_rank) + 1/(k + keyword_rank)  # k=60 (standard)

Database Schema

CREATE TABLE chunks (
    id UUID PRIMARY KEY,
    document_id UUID REFERENCES documents(id),
    content TEXT NOT NULL,
    embedding vector(1024),  -- PGVector
    content_tsvector tsvector GENERATED ALWAYS AS (
        to_tsvector('english', content)
    ) STORED,
    section_title TEXT,
    content_type TEXT,
    created_at TIMESTAMP DEFAULT NOW()
);

-- Indexes
CREATE INDEX idx_chunks_embedding ON chunks
    USING hnsw (embedding vector_cosine_ops)
    WITH (m = 16, ef_construction = 64);

CREATE INDEX idx_chunks_content_tsvector ON chunks
    USING gin (content_tsvector);

Hybrid Search Query (SQLAlchemy)

async def hybrid_search(
    query: str,
    query_embedding: list[float],
    top_k: int = 10
) -> list[Chunk]:
    FETCH_MULTIPLIER = 3  # Fetch 30 for better RRF coverage
    K = 60  # RRF smoothing constant

    # Vector search subquery
    vector_subq = (
        select(Chunk.id,
            func.row_number().over(
                order_by=Chunk.embedding.cosine_distance(query_embedding)
            ).label("vector_rank"))
        .limit(top_k * FETCH_MULTIPLIER)
        .subquery()
    )

    # Keyword search subquery
    ts_query = func.plainto_tsquery("english", query)
    keyword_subq = (
        select(Chunk.id,
            func.row_number().over(
                order_by=func.ts_rank_cd(Chunk.content_tsvector, ts_query).desc()
            ).label("keyword_rank"))
        .where(Chunk.content_tsvector.op("@@")(ts_query))
        .limit(top_k * FETCH_MULTIPLIER)
        .subquery()
    )

    # RRF fusion with FULL OUTER JOIN
    rrf_subq = (
        select(
            func.coalesce(vector_subq.c.id, keyword_subq.c.id).label("chunk_id"),
            (func.coalesce(1.0 / (K + vector_subq.c.vector_rank), 0.0) +
             func.coalesce(1.0 / (K + keyword_subq.c.keyword_rank), 0.0)
            ).label("rrf_score"))
        .select_from(vector_subq.outerjoin(keyword_subq, ..., full=True))
        .order_by("rrf_score DESC")
        .limit(top_k)
        .subquery()
    )

    return await session.execute(
        select(Chunk).join(rrf_subq, Chunk.id == rrf_subq.c.chunk_id)
    )

Common Patterns

Filtered Search

results = await hybrid_search(
    query="binary search",
    query_embedding=embedding,
    content_type_filter=["code_block"]
)

Similarity Threshold

results = await hybrid_search(query, embedding, top_k=50)
filtered = [r for r in results if (1 - r.vector_distance) >= 0.75][:10]

Multi-Query Retrieval

queries = ["machine learning", "ML algorithms", "neural networks"]
all_results = [await hybrid_search(q, embed(q)) for q in queries]
final = deduplicate_and_rerank(all_results)

Performance Tips

1. Pre-compute tsvector - 5-10x faster than to_tsvector() at query time
2. Use HNSW index - 17x faster queries than IVFFlat
3. 3x fetch multiplier - Better RRF coverage (30 results per search for top 10)
4. Iterative scan for filtered queries - Set hnsw.iterative_scan = 'relaxed_order'
5. Metadata boosting - +6% MRR with title/path matching

Redis 8 FT.HYBRID Alternative

Redis 8.4 introduced FT.HYBRID with native hybrid search support. Consider this alternative when latency is critical.

FT.HYBRID Command Syntax

FT.HYBRID index query
  VECTOR $embedding AS vec_score
  TEXT "search terms" AS text_score
  COMBINE RRF
  LIMIT 0 10
  RETURN 3 content title score

Comparison: pgvector vs Redis 8 FT.HYBRID

When to Choose pgvector

• Already using PostgreSQL - No new infrastructure, single source of truth
• Need ACID transactions - Atomic updates across search index and relational data
• Complex joins with relational data - Foreign keys, aggregations, reporting
• Large datasets - Billions of vectors with table partitioning
• Compliance requirements - Established PostgreSQL security/audit tooling

When to Choose Redis 8

• Need sub-5ms latency - Real-time autocomplete, typeahead, live recommendations
• Caching layer with search - Already using Redis for cache, add search capability
• Simpler deployment - Single binary, no extensions or complex configuration
• Ephemeral search indexes - Rebuild from source of truth is acceptable
• Memory budget allows - Dataset fits comfortably in RAM

References

Detailed Implementation Guides

External Resources

• PGVector GitHub
• HNSW Index Guide

Related Skills

• ai-native-development - Embeddings and vector concepts
• database-schema-designer - Schema design for vector search

Version: 1.2.0 | Status: Production-ready | Updated: pgvector 0.8.1

Capability Details

hybrid-search-rrf

Keywords: hybrid search, rrf, reciprocal rank fusion, vector bm25, semantic keyword search Solves:

• How do I combine vector and keyword search?
• Implement hybrid retrieval with RRF
• Merge semantic and BM25 results

semantic-search

Keywords: semantic search, vector similarity, embedding, nearest neighbor, cosine distance Solves:

• How does semantic search work?
• When to use semantic vs keyword search
• Semantic search strengths and weaknesses

keyword-search-bm25

Keywords: bm25, full-text search, tsvector, tsquery, keyword search Solves:

• How does BM25 keyword search work?
• Implement PostgreSQL full-text search
• BM25 vs semantic search trade-offs

rrf-algorithm

Keywords: rrf, reciprocal rank fusion, rank-based fusion, score normalization Solves:

• How does Reciprocal Rank Fusion work?
• Why use rank instead of scores?
• RRF smoothing constant (k parameter)

database-schema

Keywords: pgvector schema, chunk table, embedding column, tsvector, generated column Solves:

• How do I design schema for hybrid search?
• Store embeddings with vector(1024)
• Pre-compute tsvector for performance

search-query-implementation

Keywords: hybrid search query, sqlalchemy, vector distance, ts_rank_cd, full outer join Solves:

• How do I write hybrid search SQL?
• Implement RRF in SQLAlchemy
• Use fetch multiplier for better coverage

indexing-strategies

Keywords: pgvector index, hnsw, ivfflat, vector index performance, index tuning Solves:

• HNSW vs IVFFlat comparison
• Optimize vector search speed
• Scale to millions of vectors

pre-computed-tsvector

Keywords: tsvector, gin index, full-text index, pre-computed column, generated column Solves:

• Optimize keyword search performance
• 5-10x speedup with indexed tsvector

metadata-filtering

Keywords: metadata filter, faceted search, content type filter, score boosting Solves:

• Filter search by metadata
• Boost results by section title
• Pre-filter by content type

common-patterns

Keywords: filtered search, similarity threshold, multi-query retrieval, search patterns Solves:

• Filter search by content type
• Set minimum similarity threshold
• Implement multi-query retrieval

golden-dataset-testing

Keywords: golden dataset, search evaluation, pass rate, mrr, retrieval testing Solves:

• Test hybrid search quality
• Evaluate search with golden queries
• Calculate pass rate and MRR metrics