active retrieval

**Active retrieval** is the **adaptive retrieval policy where the model decides when and what to retrieve during reasoning rather than using a fixed one-shot fetch** - it aligns retrieval effort with uncertainty and task complexity. **What Is Active retrieval?** - **Definition**: Decision-driven retrieval that is triggered conditionally during generation or planning. - **Trigger Signals**: Uncertainty estimates, contradiction detection, and missing-evidence indicators. - **Control Granularity**: Can choose retrieval timing, query form, and candidate budget per step. - **System Benefit**: Avoids unnecessary retrieval on simple questions and deepens search on hard ones. **Why Active retrieval Matters** - **Efficiency**: Dynamic retrieval allocates compute where it adds the most value. - **Accuracy**: On-demand evidence gathering improves support for uncertain claims. - **Latency Balance**: Skips extra retrieval when confidence is already high. - **Robustness**: Adaptive loops better handle ambiguous or evolving questions. - **Safety**: Retrieval-on-uncertainty reduces unsupported model assertions. **How It Is Used in Practice** - **Policy Learning**: Train controllers to predict retrieval utility from intermediate states. - **Confidence Instrumentation**: Expose uncertainty metrics to drive retrieval decisions. - **Guardrails**: Set max retrieval rounds and enforce citation requirements for critical outputs. Active retrieval is **a high-value optimization for adaptive RAG pipelines** - active control improves cost-quality tradeoffs while strengthening grounded responses.