I have built a Semantic Classification engine, which semantically classify codebase. This can be used by Coding Agents go through unknown codebases, understand flow in it, search semantically, finding heavy dependencies.
What happens as the code changes? Why is this better than the agent using grep and LSP in a targetted way for the task at hand, against the current code, including the current diff?
How does one get an agent to prefer this tool to the ones it's been trained on and uses almost immediately? (ignoring instructions to use alternative code understanding tools, I've tried getting them to use LSP, but they all most always ignore those instructions)
I've built something similar and I'm using it too. In my case, it just finds function, method, or class connections, and it works surprisingly well. Then I debug visually—it's been really helpful for AI coding.
Great to hear buddy. For my case, I didn't really want to create semantic classifier. I started as a tool, that can enforce architectural rules, like "no direct DB access outside repository layer, controller should only handle req/res life cycle no massive logical works". The hard part was that to enforce rules like that, the system first has to understand what a file actually is semantically: controller, service, repository, middleware, route, and so on.
So this is how I am here.
What you say about this product?