🧠 RepoMind

A repository-aware code intelligence engine powered by call graphs, repository maps, and vector search.

RepoMind analyzes a full codebase and builds a structured understanding of it using:

• Static repository structure analysis
• Function-level call graph extraction
• Semantic code search (Qdrant)
• LLM-based reasoning over code context

Currently optimized for local LLM inference via LM Studio.

---

✨ Features

• 📁 Repository structure mapping (modules, files, importance scoring)
• 🔗 Static call graph generation (cross-file function tracking)
• 🧠 Semantic code search using embeddings
• 💬 LLM-powered chat over codebase
• ⚡ Incremental indexing support
• 📦 Local-first architecture (no cloud dependency required)

---

⚠️ Current Limitations (IMPORTANT)

RepoMind is still in early development.

Current constraints:

• ❗ Only works with LM Studio (OpenAI-compatible local API)
• ❗ No OpenAI / cloud API support yet
• ❗ Requires Qdrant running locally
• ❗ No UI (CLI-only experience)

---

🧱 Architecture Overview

RepoMind builds 3 intelligence layers:

1. Repository Map

• File & folder structure
• Module classification
• Importance scoring

2. Call Graph

• Function-level relationships
• Cross-file dependencies
• Execution flow inference
• Unresolved edges tracking

3. Vector Search Layer (Qdrant)

• Code chunk embeddings
• Semantic retrieval
• Context ranking for LLM prompts

---

📦 Project Structure

src/
│
├── chat.py              # Main Q&A interface
├── indexer.py           # Embedding + Qdrant indexing
├── repo_map.py          # Repository structure analysis
├── call_graph.py        # Static function call graph
├── search.py            # Semantic search engine
├── reranker.py          # Context reranking
├── memory.py           # Chat memory layer
├── config.py           # Local config (ignored by git)
├── config.example.py   # Example config for setup
│
├── cache/              # Generated artifacts
└── workspace/          # Cloned repositories

---

⚙️ Requirements

1. Python

• Python 3.10+

2. Install dependencies

pip install -r requirements.txt

3. install & Start Qdrant (Vector Database)

docker pull qdrant/qdrant

docker run -p 6333:6333 -p 6334:6334 -v qdrant_storage:/qdrant/storage qdrant/qdrant

---

⚙️ Configuration Setup

Step 1 — Create config file

cp src/config.example.py src/config.py

Step 2 — Configure repository

Edit only this line:

REPO_URL = "https://github.com/your-username/your-repo.git"

---

Step 3 — LM Studio setup (IMPORTANT)

RepoMind currently uses LM Studio as the only supported LLM backend.

Make sure:

• LM Studio is running
• Server is enabled
• API endpoint is:

http://localhost:1234/v1

In config.py:

LMSTUDIO_URL = "http://localhost:1234/v1"
LMSTUDIO_KEY = "lm-studio"

---

🚀 Running the System

1. Build Repository Map

python -m src.repo_map

Output:

src/cache/repo_map.json

---

2. Build Call Graph

python -m src.call_graph

Output:

src/cache/call_graph.json

---

3. Index Codebase (Qdrant)

python -m src.indexer

This step:

• Clones repo
• Chunks code
• Generates embeddings
• Stores vectors in Qdrant

---

4. Start Chat Interface

python -m src.chat

---

💬 Example Queries

Try:

Explain websocket flow step by step

Trace login request from client to server response

Where is authentication handled in the system?

How does call graph resolve cross-file functions?

---

🧠 How It Works

RepoMind combines:

• 📁 Structural understanding (repo map)
• 🔗 Execution relationships (call graph)
• 🔍 Semantic search (vector DB)
• 🧠 LLM reasoning (LM Studio)

This creates a multi-layer code reasoning system.

---

⚠️ Important Notes

• config.py is ignored in git (contains local settings)
• cache/ is auto-generated
• workspace/ contains cloned repositories
• Always run from project root
• First run may take time due to indexing

---

🔮 Roadmap

Planned improvements:

• Multi-repo support
• OpenAI / Claude backend support
• Web UI (graph + chat)
• Real-time AST-based call graph
• Incremental background indexing
• Better unresolved call resolution

---

📌 License

MIT