feat: code embeddings combined with graph structure for improved retrieval

[vitali87]

Task Description

Explore combining code embeddings with graph structural information to improve code retrieval quality, inspired by CodeFuse-CGM which achieved high rankings on SWE-bench using this approach.

The idea is to enrich the vector similarity search with graph topology signals (e.g., call graphs, import chains, class hierarchies) so that structurally related code is ranked higher even when lexically dissimilar.

Acceptance Criteria

• Research the CodeFuse-CGM approach and identify applicable techniques
• Prototype a retrieval method that combines embedding similarity with graph structure scores
• Benchmark retrieval quality against the current approach

References

• Original discussion: Run this project as MCP Server #61 (comment by @qdrddr)
• CodeFuse-CGM: https://github.com/codefuse-ai/CodeFuse-CGM/

Priority

• High
• Medium
• Low

Estimated Effort

• Small (< 2 hours)
• Medium (2-8 hours)
• Large (> 8 hours)