CodeRevU — AI-Powered GitHub PR Review

CodeRevU is designed around a fully event-driven, server-rendered Retrieval-Augmented Generation (RAG) architecture:

• •Next.js 16 App Router Platform: Coordinates client interactions, Server Actions, API routes, and real-time dashboard UI using shadcn/ui and TanStack Query.
• •Auth: Better Auth validates sessions and enables credential-less GitHub OAuth flow.
• •Webhooks: Secures incoming GitHub repository updates via HMAC-SHA256 signatures, routed to a dedicated PR event handler validated by Zod schemas.
• •Dashboard: Renders dynamic repo analytics, review histories, and active project settings.
• •Payments: Manages free/pro subscription tiers and limits using Polar.sh checkout and billing gates.
• •Inngest Background Workers: Drives reliable, durable asynchronous queue workflows with built-in retry mechanisms and fine-grained concurrency limiters.
• •indexRepo: Triggered on project registration to crawl repository files via Octokit, generate embeddings, and bulk upsert them into the vector store.
• •generateReview: Triggered on pull request updates to analyze diff files, pull context-aware matches from the vector database, and format structured code feedback.
• •Data Stores: PostgreSQL (interfaced via Prisma ORM) manages user configurations and repository states. Pinecone DB holds 768-dimensional vector embeddings for low-latency similarity queries.
• •AI/LLM Engine: Google Gemini 2.5 Flash acts as the core model for both compiling vector embeddings during ingestion and formulating targeted reviews under precise grounding prompts.

• •Indexing large repositories without blocking GitHub webhook flows.
• •Designing a RAG pipeline that retrieves semantically relevant code instead of naive file matches.
• •Handling concurrent PR events safely across multiple repositories.
• •Preventing hallucinated feedback by grounding LLM responses in real code context.

• •Pinecone-based vector search significantly improved contextual relevance over keyword-based approaches.
• •Asynchronous workflows with Inngest prevented API blocking and enabled safe concurrency.
• •Structured prompts reduced noisy or vague review output.
• •Webhook-driven architectures require idempotency and replay safety.

• •PR-level RAG instead of generic repo chat.
• •Fully automated review generation on PR open/update.
• •Structured feedback designed for real engineering teams.

• •Eliminated repetitive manual review effort for common PR patterns.
• •Enabled faster review cycles with consistent feedback quality.