Back to full roadmap
topicadvanced
Hybrid Retrieval (Dense + Sparse)
Vector search + BM25 keyword fusion — 15-25% recall improvement.
3 hours2 resources1 prereqs
Dense embeddings (semantic) miss on names, code, abbreviations, numbers. BM25 (classic IR) fills that gap.
Reciprocal Rank Fusion (RRF): combine ranks across result lists, the most consistently top-ranked item wins.
Stack: Qdrant / Weaviate / Elastic do hybrid out-of-the-box. pgvector + pg_trgm is Postgres-native.