This is what aider does, using tree sitter to extract the AST from each source file. It uses the ASTs to build a call graph. And then does a graph optimization to identify the most relevant parts of the code base, given the current state of the LLM chat.
There’s more details in this article: