Problem Description

I am building an Ai code generation workflow to go from task (ticket, description, etc) to github PR with generated code based on the codebase context. its able to get relevant codebase context, but inconsistencies with knowing which files to edit, the dependency graph, and code validation.

Stack:
LanceDB for vector embeddings
OAI for embeddings, code generation
LLamaIndex's Codesplitter for AST-based chunking

Acceptance Criteria

Be able to go from natural language description of what you want to do within the codebase, and it edits/creates/deletes the necessary files with high accuracy (similar to devin of sorts).