passage retrieval

**Passage retrieval** is the **retrieval task of finding the most relevant short text spans rather than whole documents for answering a query** - it is central to RAG because generation quality depends on precise, context-sized evidence. **What Is Passage retrieval?** - **Definition**: Search process that ranks small chunks or passages by query relevance. - **Granularity Goal**: Return evidence units that fit model context limits and preserve answer-bearing detail. - **Index Unit**: Typically uses chunked passages with metadata linking back to source documents. - **Pipeline Role**: First critical step before reranking and grounded generation. **Why Passage retrieval Matters** - **Context Efficiency**: Sending full documents wastes tokens and dilutes answer signal. - **Accuracy Impact**: Correct passage selection strongly determines factual answer quality. - **Latency Control**: Smaller units improve retrieval speed and downstream processing efficiency. - **Hallucination Reduction**: Targeted evidence lowers unsupported generation risk. - **Auditability**: Passage-level evidence supports precise citation and verification. **How It Is Used in Practice** - **Chunked Corpus Build**: Split documents into indexed passages with source and position metadata. - **Two-Stage Ranking**: Use fast retrieval followed by reranking for high-precision top-k. - **Answer Attribution**: Carry passage IDs into generation for evidence-linked outputs. Passage retrieval is **the evidence-selection core of modern RAG systems** - high-quality passage ranking is required for factual, efficient, and verifiable AI responses.