iterative-retrieval

You refine codebase context iteratively by dispatching agents to search files, evaluating relevance, and refining search patterns to build a targeted context set. This progressive retrieval method ensures subagents gather pertinent information without knowing exact file paths beforehand.

The best way to find related files is using iterative retrieval to dynamically discover patterns and expand keywords. By looping through dispatch and evaluation phases, agents can automatically locate all relevant code files for complex bugs, such as those affecting authentication and session management.

Iterative retrieval optimizes RAG pipelines by progressively fetching the most pertinent information and excluding irrelevant paths. This dynamic evaluation loop maximizes token usage accuracy, ensuring agents retrieve only the highly relevant context needed to execute subagent tasks.

Yes, you can explore codebases by employing an iterative retrieval loop that supports dynamic pattern discovery. Agents evaluate search results and expand keywords over multiple cycles, successfully identifying relevant files within specified directories without requiring exact file names.

A limitation of iterative context retrieval is its strict maximum of three cycles within the dispatch, evaluate, and refine loop. This caps the exploration depth, meaning extremely dispersed codebases might require manual intervention if relevant paths are not discovered within the three-cycle limit.