iterative retrieval

**Iterative retrieval** is the **retrieval strategy that repeatedly refines queries and candidate selection based on intermediate findings** - it improves evidence quality when initial retrieval is incomplete or noisy. **What Is Iterative retrieval?** - **Definition**: Multi-round retrieval loop where each round uses context from previous rounds. - **Refinement Signals**: Uses partial answers, uncertainty cues, or missing-entity detection. - **Stopping Criteria**: Terminates on confidence threshold, max rounds, or saturation of new evidence. - **Pipeline Role**: Bridges retrieval and reasoning for hard information needs. **Why Iterative retrieval Matters** - **Coverage Recovery**: Second or third rounds can find evidence missed by first-pass queries. - **Noise Reduction**: Later rounds can focus search space using validated intermediate facts. - **Answer Robustness**: Progressive refinement lowers chance of premature incorrect conclusions. - **Adaptivity**: System reacts dynamically to ambiguous or under-specified user input. - **Practical Accuracy**: Often improves outcomes on long-tail and multi-step questions. **How It Is Used in Practice** - **Loop Controller**: Track evidence gain and confidence at each retrieval iteration. - **Query Rewriter**: Generate focused follow-up queries from unresolved sub-questions. - **Budget Governance**: Cap rounds and compute usage to preserve latency objectives. Iterative retrieval is **a useful strategy for hard-query evidence discovery** - iterative loops trade modest extra compute for stronger retrieval completeness and answer reliability.