RAG Pipeline Design & Failure Modes

You are a retrieval engineer reviewing a RAG system for [DOMAIN] docs ([PDF/NOTION/CONFLUENCE/etc], ~[DOC_COUNT] documents). Given constraints: [EMBEDDING_MODEL], [VECTOR_DB], [CHUNK_SIZE], [TOP_K], [LLM] — produce: 1) Ingestion blueprint: chunking strategy (semantic vs fixed), metadata fields to extract, deduping, versioning 2) Query pipeline: hybrid search recipe (BM25 + vector weights), reranker choice, query rewriting prompts, and when to abstain 3) Top 8 failure modes (hallucination, stale docs, wrong chunk, ACL leaks…) with detection + mitigation 4) Grounded answer template forcing citations [1][2] and "I don't know" behavior 5) Offline eval: 20 Q/A pairs format + metrics (nDCG@10, faithfulness, latency p95) 6) Cost model: embedding $, storage, and $/1k queries at [VOLUME]/month If I paste a sample failed answer: [PASTE]. Diagnose which stage broke and propose the smallest fix first.