GOG pitches AST graphs over vector RAG

This is an interesting attack on a real weakness in code RAG: repositories are structured graphs, not loose prose, so deterministic traversal can plausibly beat embedding search on path-sensitive tasks.

• –The core bet is that a symbolic dependency graph can filter context far more aggressively than chunk-and-rank retrieval, which matters most in large repos full of red herrings.
• –The reported wins are big, but they come from an author-built benchmark maze, so the next question is whether the approach holds up on messy real-world monorepos, dynamic imports, and polyglot stacks.
• –Even if the “replace RAG” framing is too absolute, GOG points toward a strong hybrid pattern where graph traversal does candidate selection and the LLM handles synthesis.
• –The GitHub repo makes this more than a thought experiment: there is runnable benchmark code, a PDF paper, and both cloud and local-model paths for testing.
• –For AI coding tools, the most compelling angle is not raw speed alone but reducing hallucinated file paths and irrelevant context during deep code navigation.