AI Code Review System

A monorepo for automated code review using AI.

Data Flows

1. When a User Installs the Bot to a Repository

sequenceDiagram
    participant User
    participant GitHub
    participant Web as apps/web
    participant WS as services/webhook-service
    participant DB as packages/db
    participant RI as services/repo-indexer

    User->>Web: User initiates GitHub App installation
    Web->>GitHub: Redirect to GitHub OAuth
    GitHub->>Web: OAuth callback with installation token
    Note over Web,GitHub: User selects repositories to install

    GitHub->>WS: Webhook: installation created
    WS->>DB: Create installation record
    DB-->>WS: Installation created

    loop For each repository
        WS->>DB: Upsert repository
        WS->>DB: Upsert RepositoryBranch (default branch)
        WS->>DB: Create RepoIndexingJob (status: pending)
    end

    WS->>RI: Add job to repo-index queue (with branchId, jobId)
    RI->>DB: Update job status to "processing"
    RI->>DB: Update branch indexingStatus to "indexing"
    RI->>GitHub: Fetch repository files
    RI->>RI: Filter excluded files (package.json, images, binaries, etc.)
    RI->>RI: Compute file hash (SHA-256)
    RI->>RI: Chunk large files (500 lines, 50 overlap)
    RI->>DB: Create/update IndexedFile records
    RI->>DB: Create FileChunk records with embeddings
    RI->>RI: Store vectors in Pinecone
    RI->>DB: Update job status to "completed"
    RI->>DB: Update branch indexingStatus to "indexed"

2. Repository Indexing Flow (Detailed)

flowchart TD
    A[Start Indexing Job] --> B{Job already processing?}
    B -->|Yes| C[Skip - avoid duplicates]
    B -->|No| D[Update job status: processing]
    D --> E[Update branch indexingStatus: indexing]
    E --> F[Fetch files from GitHub]
    F --> G[Filter excluded files]
    G --> H{For each file}
    H --> I{File exists in DB?}
    I -->|Yes| J{Hash changed?}
    J -->|No| K[Skip - file unchanged]
    J -->|Yes| L[Delete old chunks, re-index]
    I -->|No| M[Create new IndexedFile]
    M --> N[Chunk file content]
    N --> O[Generate embeddings]
    O --> P[Create FileChunk records]
    P --> Q[Upsert vectors to Pinecone]
    Q --> R{More files?}
    R -->|Yes| H
    R -->|No| S{Any failures?}
    S -->|Yes| T[Update job status: pending, schedule retry]
    S -->|No| U[Update job status: completed]
    U --> V[Update branch indexingStatus: indexed]

3. When a User Raises a Pull Request

sequenceDiagram
    participant User
    participant GitHub
    participant WS as services/webhook-service
    participant DB as packages/db
    participant PP as services/pr-processor
    participant RI as services/repo-indexer
    participant AIR as services/ai-review-worker

    User->>GitHub: Creates/opens PR
    GitHub->>WS: Webhook: pull_request opened

    WS->>DB: Find repository
    DB-->>WS: Repository found
    WS->>PP: Add job to pr-review queue

    PP->>GitHub: Get PR details & diff
    PP->>DB: Create review record (status: pending)
    PP->>RI: Add job to pr-context queue
    PP->>PP: Add job to pr-comment (initial comment)

    PP->>GitHub: Post "processing" comment

    RI->>RI: Fetch similar code from vector store
    RI->>AIR: Add job to pr-ai-review queue

    AIR->>AIR: Generate AI code review
    AIR->>DB: Update review record (status: completed)
    AIR->>PP: Add job to pr-issues queue
    AIR->>PP: Add job to pr-comment (summary)

    PP->>GitHub: Post inline comments for each issue
    PP->>GitHub: Post summary comment

Indexing Features

File Filtering

The repo-indexer automatically excludes:

• Package files: package.json, package-lock.json, yarn.lock, pnpm-lock.yaml
• Binary assets: Images (.png, .jpg, .gif, .svg, .webp), Videos, PDFs
• Build artifacts: node_modules, dist, build, .next, .nuxt
• Minified files: .min.js, .min.css, .map
• System files: .DS_Store, Thumbs.db

Chunking Strategy

• Chunk size: 500 lines
• Overlap: 50 lines (for context continuity)
• Each chunk gets its own embedding stored in FileChunk table

Retry Mechanism

• Retryable errors: Rate limits, quota exceeded
• Max retries: 3
• Backoff: Exponential (60s, 120s, 240s)
• Failed jobs are marked as pending and retried automatically

Diff-Based Indexing (Production Optimization)

The system uses intelligent diff-based indexing to avoid full re-indexing:

Branch Create Event:

• If base branch is indexed → mark new branch as indexed (no job)
• If base branch not indexed → trigger full indexing

Push Event:

• Compare before (old commit SHA) vs after (new commit SHA)
• If identical → skip indexing
• If different → create diff-based job with baseCommitSha + headCommitSha

Worker Logic:

• Uses GitHub Compare API: GET /repos/{owner}/{repo}/compare/{base}...{head}
• Only fetch changed files (added, modified, removed)
• Skip unchanged files via hash comparison
• Delete removed files from index

Edge Cases Handled:

• Force push → fallback to full indexing
• First commit (no base) → fallback to full indexing
• Large diff (500+ files) → fallback to full indexing

What's inside?

Apps and Packages

• apps/web: a Next.js app
• apps/server: a Node.js API server
• packages/ai: AI utilities shared across services
• packages/config: shared configuration
• packages/db: Prisma database client
• packages/kafka: Kafka utilities
• packages/logger: logging utility shared across services
• packages/redis: Redis client
• packages/types: shared TypeScript types
• services/ai-review-worker: AI review worker service
• services/pr-processor: PR processor and GitHub comment service
• services/repo-indexer: Repository indexing service
• services/webhook-service: Webhook service

Each package/app is 100% TypeScript.

Utilities

• TypeScript for static type checking
• ESLint for code linting
• Prettier for code formatting
• Turborepo for build orchestration

Build

To build all apps and packages, run the following command:

pnpm build

Develop

To develop all apps and packages, run the following command:

pnpm dev

You can develop a specific package by using a filter:

pnpm dev --filter=web
pnpm dev --filter=ai-review-worker

Roadmap - Feature Parity with CodeRabbit

Already Implemented ✅

• PR webhook handling
• Repository indexing with Pinecone vector store
• AI code review generation (Gemini)
• Inline comments + summary with walkthrough, sequence diagrams, poem
• GitHub Check Run integration
• Real-time streaming (SSE)
• Dashboard with review history
• Repository management page

Todo List - Missing Features

• [ ] Chat/Conversation - Chat interface for follow-up questions about code
• [x] Auto-PR description - Automatically generate PR descriptions
• [ ] Custom review rules - Configurable rules (security, performance, best practices)
• [ ] Accept/Reject UI - Dashboard to accept AI suggestions directly
• [ ] Security-specific scanning - Dedicated vulnerability detection
• [ ] Test generation - Suggest or generate tests
• [ ] Code explanation - Ask for explanations of specific code sections
• [ ] Review automation rules - Configure when to auto-review (labels, paths, authors)
• [ ] Team/Organization management - Multi-tenant support
• [ ] Billing/Subscription - Monetization features
• [ ] API for external access - Programmatic integrations
• [x] Review analytics - Statistics on review patterns
• [ ] File deep-dive view - Detailed per-file review UI in dashboard
• [ ] Commit-level analysis - Analyze individual commits (beyond just PRs)
• [ ] Notification preferences - Granular alert settings