iterative-retrieval

Iterative retrieval progressively discovers and refines the small set of code files needed for a task. It uses a dispatch-evaluate-refine loop to rank files by relevance and stop early once sufficient context is found.

RAG-style code exploration applies a bounded dispatch-evaluate-refine loop, scoring file relevance and expanding queries with discovered patterns. It automatically learns project terminology and file patterns to focus subsequent searches up to three cycles.

Bug triage benefits from iterative retrieval's query and pattern learning, which automatically discovers project terminology like auth or jwt. It starts broad and refines searches to locate relevant source files even when exact terms are unknown.

Multi-agent code exploration workflows use iterative retrieval to retrieve and progressively refine relevant codebase context. The process ranks files by relevance, extracts context gaps, and converges after finding high-relevance files for feature implementation.

Iterative retrieval stops on convergence or after three cycles. It features early exit capabilities, halting the dispatch-evaluate-refine loop once sufficient high-relevance files are found and critical context gaps are identified.

RAG repository retrieval may return irrelevant files without iterative context refinement. Applying iterative dispatch-evaluate-refine loops ranks files by relevance score and expands queries with discovered patterns to filter out excessive context.