gersteinlab/LocAgent

[ACL 2025] Graph-guided agentic framework for code localization https://arxiv.org/abs/2503.09089

What it solves

LocAgent addresses the difficulty of code localization—the process of identifying the exact files or code snippets that need to be changed to fix a bug or implement a feature. Traditional methods often struggle to navigate complex codebases and bridge the gap between natural language problem descriptions and the actual code elements.

How it works

LocAgent parses codebases into directed heterogeneous graphs that represent code structures and their dependencies. This lightweight graph representation allows an LLM agent to perform multi-hop reasoning, navigating through the codebase's dependencies to search for and locate relevant entities more effectively than simple text search.

Who it’s for

This framework is designed for developers and researchers working on automated software maintenance, AI-driven bug fixing, and LLM-based code analysis tools.

Highlights

  • Graph-Guided Navigation: Uses heterogeneous graphs to capture code dependencies for better search accuracy.
  • Multi-Hop Reasoning: Enables agents to traverse the codebase structure to find relevant code.
  • Fine-Tuned Models: Provides specialized versions of Qwen2.5-Coder (7B and 32B) optimized for code localization.
  • Benchmark Integration: Compatible with Loc-Bench and SWE-bench for evaluating localization accuracy.

Related

  • Project
    CodeBoarding/CodeBoarding

    CodeBoarding creates visual architecture maps and documentation for codebases by combining static analysis and LLM reasoning, helping developers and AI agents understand system structure.

  • Project
    trailhq/Graft

    Graft is an open‑source TypeScript/Node.js tool that builds a local, markdown‑based knowledge graph of a codebase (summaries, key snippets, typed links) using tree‑sitter for structural parsing and optional LLM summarisation. The graph is cached in a `graft/` folder, refreshed only on file changes, and automatically wired into coding agents like Claude Code, Cursor, Codex, and Gemini. Benchmarks claim up to 46 % fewer tool calls, 42 % token savings, 60 % latency reduction, and a 12‑point correctness boost on SWE‑bench. Integration is via a single `graft init` command that writes agent‑specific instruction files; the CLI also offers search, visualization, and telemetry controls. The project is MIT‑licensed and targets faster, cheaper AI‑assisted development.

  • Project
    vercel-labs/lat.md

    lat.md is an npm‑distributed CLI that creates a markdown‑based knowledge graph for a codebase. It lets you write linked documentation in `lat.md/`, annotate source files with `@lat` comments, and validates the graph. Agents and developers can search it semantically (offline MiniLM or remote OpenAI/Vercel embeddings) and use commands like `lat check`, `lat locate`, and `lat expand` to keep design decisions, constraints, and test specs in sync with the code.

  • Project
    Egonex-AI/Understand-Anything

    A multi-agent tool that turns codebases and knowledge bases into interactive knowledge graphs, helping developers visualize architecture and business logic.

  • Project
    Cranot/roam-code

    roam‑code is a local static‑analysis CLI (and optional MCP server) that builds a symbol graph of a repository, letting AI coding agents query definitions, callers, test impact, and risk (blast radius) without any remote API calls. It provides pre‑flight checks, health summaries, verification gates, and Claude‑specific hooks, all open‑source and privacy‑preserving.