retrieval-review

Use when reviewing or auditing a retrieval / vector-index / RAG pipeline for quality before it silently feeds wrong or missing context to a generator. Triggers — /retrieval-review, 'review my RAG pipeline', 'audit this vector index', 'why is retrieval bad', 'check my hybrid search', 'review my embeddings retrieval', 'is my RRF/reranker configured right', 'audit my vector DB setup'. Runs a structured seven-axis review (eval foundation, corpus & chunking, embedding geometry, index & ANN fidelity, retrieval composition, rank fusion, reranking) in a fixed load-bearing order, grounded in BEIR/MTEB evaluation methodology, the anisotropy & alignment/uniformity geometry literature, ANN-index recall theory, and the rank-fusion (RRF) and cross-encoder reranking research. Produces a severity-ranked findings list led by the issues that make the right document structurally unretrievable — each with the smell, the offending layer, why it's a problem, the metric that proves it, and a concrete fix.