Retrieving code by "how it is done" — ProjAgent and procedural similarity for repository-level code generation

This work (ProjAgent) improves how AI code generation finds the existing code it references, from a new angle.

1What makes repository-level generation hard

Repository-level code generation must implement a target function while accounting for complex cross-file dependencies and project-specific conventions. Finding the right reference code is key, yet existing retrieval methods rely predominantly on lexical, structural, or semantic similarity.

As a result they often overlook functions that implement similar procedural logic — the best references — when identifiers or application domains differ.

2Procedural similarity as a retrieval signal

ProjAgent introduces this procedural similarity as an explicit retrieval signal. It first decomposes the target function into intermediate reasoning steps, then, for each step, uses an agentic workflow to retrieve repository functions that exhibit similar procedural behavior. The retrieved procedural context is integrated with conventional semantic retrieval to construct a richer repository context for code generation.

It also adds a conservative feedback loop that iteratively repairs generated code using compiler and static-analysis feedback.

341.14% Pass@1 on REPOCOD

Evaluated on the REPOCOD benchmark, ProjAgent achieves 41.14% Pass@1, outperforming existing retrieval-based baselines. These results demonstrate that procedural similarity is an effective, previously unexplored retrieval dimension for repository-level code generation.