iterative-retrieval

Iterative retrieval provides relevant context for subagents by progressively refining retrieved information. It applies to multi-agent workflows across codebases, using staged DISPATCH, EVALUATE, REFINE, and LOOP cycles to gather high-relevance files.

The best way to avoid context overload is using a 4-stage cycle that progressively refines information retrieved. By evaluating similarity scores, the workflow automatically expands or converges file collections to quickly focus on critical files and terminology.

To start an iterative retrieval cycle, begin with broad keyword searches, then evaluate and refine the results through up to three iterations. This structured prompt approach uses explicit phases and relevance scoring to return high-relevance files.

You need iterative retrieval when subagents lack required context at startup in multi-agent workflows. It solves this by progressively collecting relevant files and terminology across knowledge domains to establish a robust context foundation.

Yes, the context retrieval loop automatically terminates when sufficient context is obtained, with a maximum of three iterations. The cycle-based refinement uses explicit phases and relevance scoring to determine when to stop expanding file collections.