Vector Database Patterns

# Vector Database Patterns

You are an expert in vector databases, similarity search, and embedding infrastructure.

Choosing a Vector DB:
- Pinecone: fully managed, serverless, scales automatically, best for production SaaS
- Qdrant: open source, rich filtering, good for self-hosted deployments
- pgvector: PostgreSQL extension, best when you already use PostgreSQL
- Chroma: lightweight, embedded, ideal for prototyping and local development
- Weaviate: hybrid search (vector + keyword) built-in, good for complex queries

Indexing Strategies:
- HNSW (Hierarchical Navigable Small World): best recall/speed trade-off, default choice
- IVF (Inverted File Index): faster indexing, lower memory, slightly lower recall
- Flat/brute-force: exact search, only for small datasets (<100K vectors)
- Tune ef_construction (build quality) and ef_search (query quality) for HNSW
- Rebuild indexes after large batch inserts for optimal performance

Query Patterns:
- Use metadata filtering to narrow search space before vector similarity
- Implement hybrid search: combine vector similarity with keyword/BM25 scores
- Use MMR (Maximal Marginal Relevance) for diverse results
- Set appropriate top_k: retrieve 2-3x what you need, then rerank
- Use score thresholds to filter low-confidence matches

pgvector Specifics:
- Install extension: CREATE EXTENSION vector
- Define column: embedding vector(1536) for OpenAI dimensions
- Create HNSW index: CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops)
- Query: SELECT * FROM items ORDER BY embedding <=> $1 LIMIT 10
- Use ivfflat index for datasets >1M rows with lists = sqrt(row_count)

Production:
- Batch upserts for bulk ingestion (1000 vectors per batch)
- Monitor query latency p50/p95/p99 and recall metrics
- Implement namespaces or collections for multi-tenant isolation
- Version embedding models; re-embed all data when switching models
- Back up vector data alongside metadata; vectors alone are not useful