pgvector-knowledge-patch

Use pgvector iterative index scans to repeatedly search approximate nearest neighbors when combining vector search with WHERE filters, mitigating overfiltering by expanding the search scope until enough results match the criteria.

pgvector 0.7.0 and later introduces halfvec for half-precision storage, sparsevec for sparse vectors, and bit indexing, each with specific operator classes and dimensional limits for optimized distance calculations.

Apply binary quantization in PostgreSQL vector search using the binary_quantize function to compress embeddings, then use a re-ranking strategy to sort the filtered candidates by exact distance for improved recall and performance.

Yes, you can use expression indexes with HNSW in pgvector to index quantized or subvector extractions directly, allowing the database to satisfy filtered vector queries efficiently without rescanning the base table.

Tune HNSW and IVFFlat indexes in pgvector 0.8.2 by configuring indexing parameters like m and ef_construction for HNSW, and lists and probes for IVFFlat, while selecting appropriate operator classes for L1, L2, or cosine distance.

Schema migrations adding halfvec or sparsevec columns fail when exceeding dimensional limits or using incompatible operator classes, requiring you to verify vector type constraints and apply correct distance metric functions for successful indexing.