RAG AI Code Reviewer
Codebase-aware pull-request reviews via a from-scratch RAG pipeline
A GitHub App that reviews PRs against the repo's own patterns — it retrieves the top-5 most semantically similar functions so suggestions match established conventions instead of giving generic feedback.
- Full RAG pipeline from scratch — function-level AST chunking with Tree-sitter, 1536-dim embeddings in pgvector, cosine-similarity retrieval. No third-party RAG framework.
- Two-process architecture (webhook receiver + async worker) bridged by a pg-boss queue with idempotent, exactly-once processing under retries and network failure.
