similarity-search-patterns

Use a vector database client like Pinecone, Qdrant, pgvector, or Weaviate to upsert embedding vectors and query by nearest neighbors. The templates provide complete classes handling index creation, batch upserts, filtered search, and reranking.

Pinecone is a managed serverless option, Qdrant offers self-hosted search with quantization and rich filtering, and pgvector adds vector search to existing PostgreSQL databases. Choose based on scale, infrastructure, and whether you already run Postgres.

Flat indexes give exact results with O(n) search for small datasets, HNSW graph indexes reach 95-99% recall with O(log n) search for medium to large data, and IVF+PQ quantization suits very large datasets with 90-95% recall.

Use hybrid search: pgvector template combines vector similarity with PostgreSQL full-text search via ts_rank, while Weaviate offers a built-in hybrid mode with an alpha parameter blending BM25 and vector scores.

Yes, the pgvector template filters on JSONB metadata fields using SQL WHERE clauses combined with vector distance ordering. This lets you pre-filter documents by metadata before ranking by embedding similarity.

Avoid ANN indexes like HNSW when your dataset is small enough for exact flat search or when you need guaranteed 100% recall. Start with a flat index and scale up only after measuring recall and latency requirements.