postgres-vectors

Vector similarity in Postgres uses pgvector to store 384-dimensional embeddings and query them with cosine distance operators. Create a table with vector(384) columns, then use the <=> operator to find similar embeddings by computing cosine distance across stored vectors.

Semantic search on embeddings scales with two index strategies: HNSW for hierarchical navigation with tuned parameters, or IVFFlat for inverted flat clustering with quantization. Choose HNSW for high recall or IVFFlat for lower memory overhead on large 384-dimensional vector sets.

Yes, deduplication by source_text prevents redundant embeddings in your table. Apply unique constraints or filter queries on source_text to ensure each embedding stores only once, reducing storage and query overhead.

Batch cosine similarity queries use the <=> operator to compute distances between a query embedding and all stored vectors in a single SQL statement. Order results by distance and limit to retrieve the top matches efficiently.

No, pgvector requires the vector extension installed in the extensions schema and a column defined as vector(384) to store 384-dimensional embeddings. Create the extension and table structure before inserting or querying embedding data.

HNSW provides better recall and query speed through hierarchical graph navigation, while IVFFlat clusters vectors for lower memory use and faster index builds. Choose HNSW for accuracy-critical search or IVFFlat for cost-constrained large-scale deployments.