TIL: Jun 12

TIL: When optimizing HNSW vector indexes for RAG, scaling down embedding dimensions (e.g., from 3072 to 1536) can significantly improve index stability and performance while maintaining semantic integrity via Matryoshka-style truncation. Always verify index utilization using EXPLAIN ANALYZE to ensure the query planner is actually traversing the HNSW graph rather than falling back to sequential scans.