Match each retrieval technique to the specific failure it is designed to fix

Naive RAG embeds the query, finds the top-k nearest chunks, and stops. Each technique here patches a different way that pipeline lets you down. **Hybrid retrieval** adds keyword (sparse) search alongside dense embeddings. Dense vectors capture meaning but blur exact tokens — a part number, an error code, a rare acronym. Sparse search matches those literally, so combining the two covers both. Reranking fixes precision. The first stage may pull good candidates but order them poorly, so a cross-encoder rescores the top results by reading query and chunk together, pushing the truly relevant ones up. Parent-document retrieval fixes fragmentation. Tiny chunks match precisely but are too small to answer from, so you match on the small chunk and feed the larger parent to the LLM. HyDE fixes query-document mismatch. A short question looks nothing like a long answer passage, so you have the LLM draft a fake answer and retrieve against that instead. MMR fixes redundancy. If the top results are near-duplicates, MMR swaps some for diverse chunks so the context budget isn't wasted on repetition.

Each technique targets one specific failure of naive top-k dense retrieval. The skill is diagnosing which failure you actually have before reaching for a fix. Failures of matching come first — the right text never scores high enough to be retrieved. Hybrid retrieval handles the case where dense embeddings smear exact tokens like IDs, error codes, or acronyms. Bolting BM25 (sparse) onto the dense search recovers literal matches, usually fused with reciprocal rank fusion. HyDE handles a subtler matching gap. A terse question embeds far from the verbose, declarative passage that answers it. So you have the model hallucinate a plausible answer and retrieve against that hypothetical document's embedding, which sits closer to the real answers. The rest are failures of the result set — the right candidates are present but the set is wrong in shape. Cross-encoder reranking fixes ordering: first-stage retrieval optimizes recall cheaply, then a slower cross-encoder reads each query-chunk pair jointly and rescores for precision, so distractors fall out of the top slots. Parent-document retrieval fixes granularity: you index small chunks for precise matching but return the larger parent so the LLM has enough surrounding context to answer. MMR fixes redundancy: it penalizes a candidate's similarity to already-selected results, so the top-k carries diverse information instead of five paraphrases of one passage. Match wrong and you waste effort — reranking can't rescue a passage that recall never retrieved.