PR Buddy

An AI-powered review companion for pull requests. Authors train agents with context about their PR, and reviewers can then ask questions — getting answers grounded in author knowledge, code diffs, and team context.

How It Works

sequenceDiagram
    participant Author
    participant PRBuddy as PR Buddy
    participant Weaviate as Knowledge Base<br/>(Weaviate)
    participant GitHub
    participant Unblocked

    Note over PRBuddy,Weaviate: Session Creation
    PRBuddy->>GitHub: Fetch diff, description, comments
    GitHub-->>PRBuddy: PR content
    PRBuddy->>Weaviate: Auto-index diff chunks,<br/>description, comments

    Note over Author,Weaviate: Author Training
    Author->>PRBuddy: "Edgar likes this approach"
    PRBuddy->>Weaviate: Index author_explanation<br/>(preserves entity names)
    PRBuddy-->>Author: Got it - Edgar approves<br/>the explicit access approach.

    opt Research (team context, history, Slack)
        Author->>PRBuddy: "What did the team discuss?"
        PRBuddy->>Unblocked: Search team context,<br/>Slack, historical decisions
        Unblocked-->>PRBuddy: Relevant discussions
        PRBuddy->>Weaviate: Index findings
        PRBuddy-->>Author: Team discussed via Envoy...
    end

Loading

sequenceDiagram
    participant Reviewer
    participant PRBuddy as PR Buddy
    participant Weaviate as Knowledge Base<br/>(Weaviate)
    participant Unblocked

    Note over Reviewer,Weaviate: Reviewer Q&A (auto-injection)
    Reviewer->>PRBuddy: "Is Edgar ok with this?"
    PRBuddy->>Weaviate: 3 parallel queries:<br/>hybrid + BM25 + vector
    Weaviate-->>PRBuddy: Matching docs<br/>(author explanations, diff, etc.)
    Note over PRBuddy: Deduplicate, rank,<br/>inject as system context
    PRBuddy-->>Reviewer: Edgar reviewed and approves.<br/>He likes the explicit single-service access.

    opt Research (when KB lacks context)
        Reviewer->>PRBuddy: "How long has this been requested?"
        PRBuddy->>Unblocked: Search Slack, Jira,<br/>historical PRs
        Unblocked-->>PRBuddy: Timeline + discussions
        PRBuddy->>Weaviate: Index findings
        PRBuddy-->>Reviewer: Requested for ~6 months,<br/>tracked in INFRA-432.
    end

Loading

Architecture

graph TB
    subgraph Frontend
        WEB[Web UI<br/>Voice + Text]
        EXT[Chrome Extension<br/>Side Panel + Diff Selection]
    end

    subgraph Server["FastAPI Server"]
        WS[WebSocket Handler]
        TEXT[Text Session]
        PIPE[Pipeline Session<br/>STT + Agent + TTS]
        CI[Context Injection<br/>Multi-query RAG]
    end

    subgraph Agents["Agent System (YAML-configured)"]
        AT[AuthorTraining]
        RQ[ReviewerQA]
        RE[Research]
        AT <-->|handoff| RE
        RQ <-->|handoff| RE
    end

    subgraph Storage
        WV[(Weaviate<br/>Vector DB)]
    end

    subgraph External
        GH[GitHub API]
        UB[Unblocked MCP<br/>Team Knowledge]
    end

    WEB --> WS
    EXT --> WS
    WS --> TEXT
    WS --> PIPE
    TEXT --> CI
    CI --> WV
    TEXT --> Agents
    PIPE --> Agents
    Agents --> WV
    RE --> GH
    RE --> UB

Loading

Key Features

• Auto-indexed PR context — Diff, description, and comments are fetched from GitHub and indexed into Weaviate on session creation. Agents start with full context immediately.
• Auto-context injection — Before each reviewer question, 3 parallel RAG queries (hybrid, BM25, vector) surface relevant author knowledge automatically. The agent doesn't need to decide to search.
• Entity-aware indexing — Author explanations preserve people, team, and system names for reliable keyword matching (e.g., "Edgar" stays "Edgar", not "a stakeholder").
• Voice + Text modes — Text chat or voice (Whisper STT + Polly/OpenAI TTS).
• Diff selection — Chrome extension lets reviewers highlight code and ask questions with that context.
• Research via Unblocked MCP — Agents can search Slack, Jira, historical PRs, and team discussions.

Quick Start

# Setup
make setup

# Edit .env and add your OPENAI_API_KEY

# Start Weaviate + dev server
make dev

Then open http://localhost:8000 in your browser.

Chrome Extension

Load the extension for the side panel experience on GitHub PRs:

1. Go to chrome://extensions
2. Enable Developer mode
3. Click Load unpacked and select the prbuddy-extension/ directory

Voice Mode Requirements

For voice modes (Pipeline), you need:

• ffmpeg — brew install ffmpeg (macOS) or sudo apt install ffmpeg (Linux)
• OPENAI_API_KEY — For Whisper STT
• AWS credentials or ELEVENLABS_API_KEY — For TTS

Session Modes

Mode	Description	Latency
Text	Chat-based Q&A	Instant
Pipeline	Whisper STT + Agent + TTS	~2-3s

Configuration

Agent behaviors are configured via YAML files in config/agents/:

Directory	Agents	Purpose
common/	Research	Context gathering from GitHub + Unblocked
author/	AuthorTraining, AuthorEngagement	Author interview and knowledge capture
reviewer/	ReviewerQA	Reviewer Q&A with citations

Development

make setup    # Install dependencies
make dev      # Start Weaviate + server (hot reload)
make test     # Run unit tests (81 tests)
make eval     # Run evals (requires Weaviate + OPENAI_API_KEY)
make clean    # Clean up generated files

Evals

The evals/ directory contains end-to-end scenarios that verify author-to-reviewer knowledge sharing:

Scenario	Tests
edgar_approval	Author says "Edgar likes this" — reviewer can find it with 5 phrasings
technical_decision	Author explains tradeoff — reviewer finds the reasoning
indirect_reference	Author provides context — reviewer finds it with different wording

make eval  # Requires Weaviate running + OPENAI_API_KEY

License

MIT