← Back to Dashboard
1. Embedding Fundamentals2. Hybrid Retrieval Strategies
Hybrid Retrieval Strategies
📚 Embeddings and Vector Search⏱ 10 min⭐ 75 XP
Dense + Keyword Works Better in Production
Hybrid retrieval combines semantic vector search with lexical (keyword/BM25) matching. Dense vectors understand meaning; lexical search nails exact tokens. Production corpora need both.
Where Each Approach Wins
| Query | Dense-only | Lexical-only | Hybrid |
|---|---|---|---|
| "how do I revoke a leaver's access" | ✓ finds "offboarding IAM checklist" | � - no keyword overlap | ✓ |
| "error INC-4402 in payment-svc" | � - IDs embed poorly | ✓ exact match on INC-4402 | ✓ |
| "policy SEC-17 exceptions" | � - may return SEC-12, SEC-19 | ✓ exact identifier | ✓ |
Bedrock Knowledge Bases supports hybrid search natively - you can set retrieval type to HYBRID (or let it default per store capability) when calling Retrieve/RetrieveAndGenerate.
The Two-Stage Pattern: Retrieve Wide, Rerank Narrow
- Stage 1 - candidate retrieval: pull a generous top-k (e.g. 25-50) using hybrid search. Optimised for recall.
- Stage 2 - reranking: score each candidate against the query with a dedicated reranker (e.g. Cohere Rerank or Amazon Rerank on Bedrock) and keep the best 3-8. Optimised for precision.
Rerankers are cross-encoders - they read query and passage together, so they judge relevance far better than raw vector distance, at a per-pair compute cost you only pay for candidates.
Tuning heuristic: if answers miss obviously relevant documents → raise stage-1 k or fix chunking (recall problem). If answers cite plausible-but-wrong passages → add or strengthen reranking (precision problem).
🧪 Knowledge Check
Press 1-4 to select1 of 2
Why use hybrid retrieval instead of dense-only?
To increase IAM permissions
To capture both semantic meaning and exact lexical signals
To disable reranking
To remove metadata