iterative-retrieval

Iterative retrieval progressively surfaces high-relevance files for subagents without overloading token limits. It uses a four-phase DISPATCH, EVALUATE, REFINE, and LOOP cycle to dynamically refine codebase context based on relevance scoring.

Refining subagent context in multi-agent workflows is best handled by an iterative loop that evaluates retrieved results and adapts queries. This context-aware refinement progressively narrows down high-relevance files across multiple cycles.

To start iterative codebase exploration, provide an initial broad query. The system dispatches the query, evaluates the retrieved results, refines the search based on relevance scoring, and loops until high-relevance files are surfaced.

Yes, progressive context refinement supports a configurable maximum cycle limit. The four-phase loop stops after reaching the maximum cycles or when the evaluation phase determines the retrieved codebase context is sufficiently relevant.

Subagents retrieve irrelevant files when exploring a large codebase due to limited initial context. Applying iterative retrieval solves this by dynamically refining queries based on evaluation scoring to progressively surface only high-relevance files.

Iterative retrieval works effectively for token-limited scenarios by progressively refining codebase context. It reduces token usage by adapting queries and retrieving only high-relevance files instead of loading entire codebases into subagent prompts.