RAG versus agentic search: retrieval research

When an agent needs context from a large multi-repo codebase, should it query a pre-built index or go and search the code live? I worked through that question for Openprovider's codebase.

I compiled the published evidence on both approaches, then applied it to the real shape of the codebase instead of a generic benchmark. The recommendation was a hybrid with three tiers doing different jobs: live search for local code, a code graph for relationships across repositories, and vector retrieval for documents that aren't code.

The output was a written study and a recommendation, grounded in measured data about the codebase rather than assumptions about it. The code-graph part of that picture is a service a teammate built, which I integrate with.