Hybrid search

Hybrid search mixes keyword search (exact terms, IDs, jargon) with vector/semantic search (paraphrases, meaning). A merger or re-ranker blends the lists.

Pure embeddings miss SKUs, error codes, and rare proper nouns. Pure keywords miss paraphrases. Real corpora need both.

1. Run BM25/keyword retrieval
2. Run dense embedding retrieval
3. Fuse scores (e.g., reciprocal rank fusion)
4. Optional cross-encoder re-rank
5. Send top chunks to the generator

Looking up “invoice #48291” (keyword) vs “how do I dispute a late fee?” (semantic).

List 3 queries for your docs that need exact match and 3 that need paraphrase match.

⚠️ Myth: Bigger embedding models remove the need for keywords.

✓ Reality: Exact tokens still win for IDs and legalese pins.

⚠️ Myth: Hybrid always costs too much.

✓ Reality: Often cheaper than wrong answers and support tickets.

• Course 09 RAG + chunking chapters
• Provider search/RAG docs (cite specifically)
• Classic BM25 IR background (Manning *Introduction to IR* — verify)