Better retrieval: hybrid search, reranking, and RAG failure modes

Naive vector search is a baseline. Production RAG layers on precision techniques:

• Hybrid search: combine semantic (vector) with keyword/BM25 search. Vectors catch meaning; keywords catch exact terms, IDs, and rare jargon vectors miss.
• Reranking: retrieve a broad top-50, then a cross-encoder reranker scores each against the query and keeps the best 5. Big precision win for modest cost.
• Query rewriting / expansion: use an LLM to clarify or expand the query, or decompose multi-part questions before retrieval.
• Metadata filtering: restrict search by date, source, or user permissions (critical for access control).

The failure modes interviewers probe

• Retrieved but ignored: model has the context but answers from priors — instruct 'answer only from context; if absent, say you don't know.'
• Right chunk missing: fix chunking/hybrid/reranking.
• Stale index: source changed but you didn't re-index.
• Conflicting sources: retrieval returns contradictory chunks — need dedup/recency/authority handling.
• Injection via documents: a malicious doc contains 'ignore your instructions' — treat retrieved text as untrusted.