BM25 (training-free keyword search using word rarity, repetition, and document length) + vector embeddings (semantic similarity) create superior search systems than either approach alone. The technologies are complementary, not competing. ## Key Insight Vector search alone misses exact keyword matches that BM25 handles naturally. BM25 alone misses semantic similarity that vectors capture. Hybrid systems leverage both: BM25 for precision on exact matches, vectors for semantic understanding. This validates a broader pattern: new technologies don't always obsolete older approaches—understanding when to combine them creates better systems than treating them as alternatives. ## Cross-Domain Connections - [[Unix Search Supremacy Over RAG]] - Simple search tools often outperform complex pipelines - [[Agentic Search Over RAG for Code]] - Real-world validation of search over embeddings - [[Local Knowledge Index Efficiency]] - Two-stage architecture (filter then process) efficiency gains - [[Simplicity Enables Workflow]] - Combining simple tools beats complex monoliths ## Source - [[Vector search is not always the answer]] — Avi Chawla, X/Twitter, February 2026 — https://x.com/_avichawla/status/2020747017258217808/?s=12&rw_tt_thread=True