I am attempting to automate the PR review process using locally running LLMs. I provide the LLM with the diff and the full file for context. However, the response usually consists of code explanations and general suggestions (e.g., logging, security) rather than an in-depth review covering complexity, alternative logic, or better approaches. Could this be an issue with the prompt? Additionally, would using RAG to index the entire codebase help ensure the review considers the complete context? Or is there a better approach? Here is the prompt I am using: https://gist.github.com/alwin-augustin-dev/a6448f53a7577e8e1614bee361aacec0
Ask HN: How to Improve LLM-Based PR Reviews for Deeper Code Analysis? | Heykuki News