Hybrid Search for Technical Docs: Why BM25 + Embeddings Beats Either Alone
2/15/2025
When users ask domain questions, exact keywords matter (part numbers, acronyms) and meaning matters (synonyms, context). Hybrid search unifies both.
Recipe we like:
- Index with BM25 for exact matches.
- Compute embeddings for semantic search.
- Merge & dedupe results, then re‑rank (e.g., cross‑encoder).
- Chunk by structure (sections/headings), not fixed tokens.
This approach powered our WeldCraft case study where users found answers 3× faster and support tickets dropped 85%.