I built this after losing hours to the same frustrating loop: ask my AI assistant an architecture question, watch it grep files, get a confident but wrong answer. The breaking point was a real bug. I asked "what handles auth token expiry in this app?" — it found the wrong file, explained the wrong flow, and I shipped a fix to the wrong place. The problem isn't the model. It's that LLMs navigate code lexically — they see files, not structure. No concept of who calls who, which class inherits what, or how a request actually flows. My first attempt was a simple symbol index. Better, but still missed connections. Then I added a call graph. Then I realized the search was broken for semantic queries — "what builds the 404 payload" would never match `core_exception_handler` lexically. That's when I added dense embeddings trained specifically on code (potion-code-16M, 16M params, runs CPU-only). Combined with BM25 via RRF fusion, it finally worked. 32 MCP tools later — context lookup, blast radius, hierarchy traversal, cycle detection, shortest path — it's the tool I actually wanted when I started. Happy to answer anything about the graph extraction, the embedding approach, or why I built it in Rust.

codesynapse

Code graph as MCP tools makes your AI stops hallucinating

Code graph as MCP tools makes your AI stops hallucinating