RepoMind 🔍

Context-Aware AI Code Review — AMD Developer Hackathon 2026

Multi-agent system that learns your repository first (AST + 3D knowledge graph), then runs context-aware code review on your PRs.

> Note: This application is designed to run locally for privacy and performance. Follow the installation guide below to run the demo.

---

🎯 Key features

1. AST analysis with tree-sitter — extracts symbols, imports, real dependencies
2. Interactive 3D knowledge graph — Three.js, force-directed, visualizes architecture
3. Automatic project profile — architecture, patterns, conventions (via LLM)
4. Multi-query retrieval — search by file, symbol, and semantics
5. Reviewer with few-shot — examples of good vs. bad justification
6. Adversarial Critic — filters generic issues before the final report
7. Privacy first — local Ollama, your code never leaves your machine
8. Secure secrets — centralized config, automatic masking in logs

---

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│  INDEXER (runs once per repo)                       │
│  Files → AST (tree-sitter) → Knowledge Graph        │
│              ↓                                      │
│       Project Profile (LLM)                         │
│              ↓                                      │
│       Vector Store (ChromaDB)                       │
│              ↓                                      │
│       3D Graph (Three.js)                           │
└─────────────────────────────────────────────────────┘

┌─────────────────────────────────────────────────────┐
│  REVIEW PIPELINE (LangGraph)                        │
│  Diff → Parse → Contextualize → Review              │
│                                     ↓               │
│                                  Critic             │
│                                     ↓               │
│                                  Report             │
│                                                     │
│  Context files GLOW in the 3D graph in real time    │
│  while the pipeline runs                            │
└─────────────────────────────────────────────────────┘

---

🚀 Quick Start (Windows)

Prerequisites

• Python 3.11+
• Node.js 18+
• Ollama installed

Automatic setup

From the project folder:

setup.bat

Does everything: creates the venv, installs deps, pulls the Ollama model, validates.

Run

:: Terminal 1
run_backend.bat

:: Terminal 2
run_frontend.bat

Open http://localhost:5173

Build as EXE

To produce a single RepoMind.exe (frontend + backend bundled):

build.bat

Output: dist\RepoMind.exe (~500MB). See BUILD_EXE.md for details.

---

🔐 Credentials security

API keys are NEVER stored in code. Everything goes through .env:

backend/.env             ← your credentials (DO NOT commit)
backend/.env.example     ← template (safe to commit)
backend/config.py        ← loads and validates secrets
.gitignore               ← protects .env

The backend:

• Validates config on startup and fails fast if anything is missing
• Automatically masks any credential that shows up in logs
• The /config endpoint returns public info (provider/model) without keys
• WebSockets sanitize messages before sending them to the frontend

Switching providers (Ollama → AMD Cloud → Groq) is just editing .env. Zero code changes.

---

🎨 Interface

3D Knowledge Graph (Three.js):

• Nodes = files, size proportional to LOC
• Colors by language
• Edges = imports
• Drag to rotate, scroll to zoom, click nodes to inspect
• During review, context files pulse in red showing exactly what the RAG retrieved

Live pipeline visualizer:

• Each agent lights up green with timing
• Shows queries used, retrieved context, issues raised vs. filtered out by the Critic

---

📁 Structure

repomind/
├── setup.bat              ← automatic Windows setup
├── run_backend.bat        ← starts the backend
├── run_frontend.bat       ← starts the frontend
├── README.md
├── .gitignore             ← protects secrets
│
├── backend/
│   ├── config.py          ← config + secret sanitization
│   ├── agents/
│   │   ├── llm_client.py
│   │   ├── code_analyzer.py    ← AST via tree-sitter
│   │   ├── indexer.py          ← + serialize_graph_for_viz
│   │   ├── diff_parser.py
│   │   ├── contextualizer.py
│   │   ├── reviewer.py
│   │   └── pipeline.py
│   ├── api/main.py        ← + /config + /repos/{id}/graph
│   ├── .env.example
│   └── requirements.txt
│
├── frontend/
│   ├── package.json       ← + three
│   ├── vite.config.js
│   ├── index.html
│   └── src/
│       ├── main.jsx
│       ├── App.jsx
│       └── components/
│           └── GraphViz3D.jsx   ← Three.js graph
│
├── scripts/
│   ├── validate_setup.py
│   └── test_e2e.py
│
└── sample_diffs/

---

🔧 Switching provider

Edit backend/.env:

Ollama (default, local)

AMD_BASE_URL=http://localhost:11434/v1
AMD_API_KEY=ollama
AMD_MODEL=qwen2.5-coder:14b

AMD Developer Cloud (when credits arrive)

AMD_BASE_URL=https://api.amd.com/v1
AMD_API_KEY=your_amd_key
AMD_MODEL=Qwen/Qwen2.5-Coder-32B-Instruct

Groq (free, fast)

AMD_BASE_URL=https://api.groq.com/openai/v1
AMD_API_KEY=gsk_...
AMD_MODEL=qwen-2.5-coder-32b

---

🎬 Demo flow

1. Point it at a real Python project (FastAPI/Django/anything)
2. Live indexing → 3D graph appears, profile extracted
3. Paste a diff that violates a project-specific pattern
4. Streaming pipeline → each agent lights up, context is retrieved
5. Relevant files pulse in the 3D graph ← this is the "wow moment"
6. Result: issues with project-specific justification

---

⚙️ Stack

Component	Tech
Orchestration	LangGraph
LLM	Qwen2.5-Coder via Ollama / AMD ROCm
AST	tree-sitter
Embeddings	BAAI/bge-small-en-v1.5 (local CPU)
Vector Store	ChromaDB
Knowledge Graph	NetworkX + Three.js
Cache	DiskCache
Backend	FastAPI + WebSockets
Frontend	React + Vite + Three.js

---

🐛 Troubleshooting

"AMD_API_KEY not configured" → cd backend && copy .env.example .env

"Could not connect to localhost:11434" → Ollama isn't running. Open the app (tray icon) or run ollama serve.

"Model not installed" → ollama pull qwen2.5-coder:14b

Slow pipeline → Check Task Manager to see if the GPU is actually being used. If it falls back to CPU, reinstall NVIDIA drivers. Or use a smaller model: ollama pull qwen2.5-coder:7b

3D graph doesn't show up → Index a repo first (step 01 in the UI). Then select it under "REPOS".