A reproducible experiment comparing traditional vector-based RAG with Agentic Search (context trees) for code retrieval tasks.
Most coding agents use vector embeddings and similarity search to retrieve code context. This works well for natural language but fails for code because code is a graph, not a "bag of words."
This experiment validates that context trees with agentic search dramatically outperform traditional RAG for code understanding tasks.
Key Results:
- 🚀 99% fewer tokens used per query
- ✅ 2× better retrieval accuracy (IoU score)
- 🎯 2.2× higher precision in finding relevant files
- 📊 Detailed Results & Analysis - Complete experimental findings with visualizations
- 🔬 Reproduction Guide - Step-by-step instructions to reproduce the experiment
- 🌐 ByteRover - The agentic search tool used in this experiment
- 📝 Blog Post - (Coming soon) Full write-up of our findings
MIT License