rag

# Retrieval-Augmented Generation ## Purpose Build a RAG system whose answers are grounded in retrieved evidence, and be able to tell — when an answer is wrong — whether the retriever failed to find the right document or the generator failed to use it. ## When to Use - Building question answering over a document corpus. - A RAG system that returns confident, wrong answers. - Choosing chunking, embedding, and retrieval strategy. - Adding citation and grounding to a generative feature. ## Capabilities - Chunking strategies and their trade-offs. - Embedding selection and hybrid (dense + sparse) retrieval. - Reranking and query rewriting. - Grounding, citation, and refusal when evidence is absent. - Component-wise evaluation. ## Inputs - The corpus: its size, structure, and update frequency. - The query patterns: factual lookup, comparison, summarization, multi-hop. - Accuracy requirements and the cost of a wrong answer. ## Outputs - A retrieval pipeline with a measured recall figure. - Answers with citations to the retrieved passages. - Separate evaluation of retrieval quality and generation quality. ## Workflow 1. **Evaluate retrieval separately** — Before touching the prompt. If the correct passage is not in the top-k, no amount of prompt engineering will produce a correct answer. Measure recall@k first. 2. **Chunk on semantic boundaries** — Sections, paragraphs, or logical units. Fixed-size chunking splits a table in half and produces two useless chunks. Add overla...