RepoGraph

🗒️ Description

RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph — a peer-reviewed paper (ICLR 2025, arXiv:2410.14684). It’s the citable academic anchor for the thesis in Structural Retrieval for Code: that a structural graph beats flat retrieval for repository-level coding tasks.

RepoGraph is a plug-and-play, framework-agnostic module — you drop it into an existing SWE agent and it adds a code-graph retrieval step. That’s what makes it interesting beyond a benchmark number: it’s designed to bolt onto a harness, not replace it.

🧩 How it works

• Graph construction — uses tree-sitter to parse the repo into ASTs. Every node is a line of code, classified as a definition or a reference. Edges are relational: E_invoke connects a definition to where it’s used, E_contain nests definitions (class → method).
• Retrieval = k-hop ego-graphs — given a search term, it returns the k-hop neighbourhood centred on that node. This is graph traversal, explicitly not embedding similarity. Defaults to 1-hop (avg ~11.6 nodes / 37.1 edges); 2-hop tested (~54.5 / 89.9).
• The retrieved subgraph becomes structured context for the downstream agent — so the model sees what connects to what, not a pile of similar-looking chunks.

📒 Results (self-reported — read the caveats)

Plugged into four systems spanning two paradigms — RAG, Agentless, AutoCodeRover, SWE-agent — on SWE-bench-Lite:

• +32.8% average relative improvement; Agentless + RepoGraph resolved 29.67% of issues (SOTA among open-source at the time).

⚠️ Flags for honesty (from the verification pass):

• The 32.8% average is inflated by a weak RAG baseline (one system showed ~99.6% relative gain off a tiny base). Read it as a self-reported average relative improvement, not a clean across-the-board win.
• “SOTA among open-source” was true Oct 2024 and is now stale versus 2026 SWE-bench leaderboards.
• SWE-bench-Lite is bug-repair; results may not transfer to line completion (where GrepRAG argues the opposite — see below).

⚔️ Tension with the grep camp

RepoGraph says “graph helps, +32.8%”. GrepRAG says “index-free grep matches or beats graph/vector RAG” on repo-level completion. The likely reconciliation is task type: graphs pay off on multi-file bug repair / cross-subsystem edits (RepoGraph’s benchmark), while agentic grep wins on local completion. This is the core “match retrieval to task” lesson in Structural Retrieval for Code.

📖 Further reading / watching

• Paper: RepoGraph, arXiv:2410.14684 (ICLR 2025)
• Structural Retrieval for Code — the lane taxonomy this anchors
• GrepRAG — the opposing “grep is enough” result
• Codebase Memory MCP · Graphify — production tree-sitter graph tools in the same lane
• Progressive Disclosure · Context Engineering — the underlying context-shaping principle

---

Template: knowledge_note_info