Code-Aware RAG Tool for Developers Seeks Feedback

Code-Aware RAG Tool — Looking for Developer Feedback

A new RAG tool is in development focused on understanding codebases, rather than treating code as simple text, aiming to provide more accurate and relevant code snippets in response to queries. The tool leverages Abstract Syntax Tree (AST) parsing and dependency graph expansion to achieve this.

Why This Matters

Traditional RAG systems often struggle with code because semantic similarity based on embeddings can miss crucial relationships between functions and calls. This leads to irrelevant or incomplete code snippets being returned, increasing developer debugging time and potentially introducing errors; a failed code suggestion can cost developers hours of rework. This new approach prioritizes structural understanding of code to mitigate these issues.

Key Insights

• AST-based chunking with Tree-sitter: Uses Tree-sitter for parsing Python, JavaScript, and TypeScript.
• Dependency Graph Expansion: Builds a dynamic graph of code dependencies to retrieve connected code paths.
• Backend-Agnostic Vector Store: Enables flexibility in storage without requiring code changes.

Practical Applications

• Codebase Search: A large software company could use this to quickly find all functions that call a specific API, including those in dependent modules.
• Pitfall: Relying solely on semantic similarity can return code snippets that look similar but are semantically unrelated, leading to incorrect implementations.

References:

• https://dev.to/demon_slayer_3e2c6835f1b2/rag-for-developers-built-for-code-not-just-text-review-requested-10d