Phase 06 · RAG - retrieval-augmented generation

Hybrid search & reranking

Combine semantic search with old-school keyword search (hybrid), then have a smarter model re-order the candidates (reranking) - the two standard upgrades when plain RAG plateaus.

In plain terms

Embeddings are great at meaning but weirdly bad at exact codes: search 'error TS-4401' and semantic search returns pages about errors, missing the one page containing TS-4401. Keyword search nails codes but misses synonyms. Hybrid runs both and merges. Reranking then adds a picky second interviewer: retrieval shortlists 25 candidates fast; the reranker reads each one carefully against the query and reorders the top picks.

Why it matters

Product names, SKUs, legal clause numbers, function names - real corpora are full of exact strings embeddings fumble. And bi-encoder retrieval (query and doc embedded separately) is inherently cruder than a cross-encoder that reads them together. These two fixes buy the biggest retrieval gains per hour of work.

How it works

Hybrid: run vector search + BM25 (keyword ranking) in parallel; merge with Reciprocal Rank Fusion (RRF - combines rank positions, no score calibration needed). Most vector DBs ship this as a flag. Rerank: over-retrieve (top 25-50), send each (query, chunk) pair to a reranker (Cohere Rerank API or an open cross-encoder), keep the top 3-5 for the prompt. Adds ~100-300ms and a small cost.

When you use it

After measuring plain RAG and finding retrieval misses - especially exact-term queries. Not on day one; measure first.

Common mistakes

  • Adding both upgrades before having a retrieval eval, so you can't tell if they helped.
  • Reranking only 5 candidates - the point is giving the reranker a WIDE net to fix.
  • Merging keyword and vector scores by raw addition (incomparable scales - use RRF).

Best practices

  • Keep the golden-set retrieval eval from Phase 04; measure hit-rate before/after each upgrade.
  • Typical winning stack: hybrid retrieve 30 -> rerank -> top 4 to the model.

Try it yourself

Add 10 exact-term queries (codes, names) to your golden set. Measure vector-only vs hybrid hit-rate. Then bolt on a reranker and measure again - three numbers, full story.

Resources

Related topics

The RAG pipeline, end to end Wiring ingestion and query-time retrieval into one system, where overall quality is the product of every stage's quality. Vector databases Databases purpose-built to store millions of embeddings and answer 'find the K nearest vectors to this one' in milliseconds. Evaluating & debugging RAG Scoring retrieval and generation separately - did we fetch the right chunks? did we answer faithfully from them? - because a single 'is it good?' hides which half is broken.