Ask questions about any codebase using AI-powered semantic search. This tool uses Retrieval-Augmented Generation (RAG) to provide accurate, context-aware answers about your code.

• 🔍 Semantic Code Search - Find relevant code using natural language
• 🤖 AI-Powered Answers - Get explanations using CodeLlama
• 📦 100% Local - No cloud calls, your code stays private
• 🚀 Fast Indexing - FAISS vector search for instant retrieval
• 💻 Clean UI - Beautiful Streamlit interface
• 🔌 GitHub Integration - Index any public repository

• "How does authentication work in this codebase?"
• "Explain the database connection logic"
• "What does the UserService class do?"
• "Show me all API endpoints"
• "Where is error handling implemented?"

1. Python 3.8+
2. Ollama - Download here

1. Clone the repository:

git clone https://github.com/harshitak4/codebase-rag.git
cd codebase-rag

2. Install dependencies:

pip install -r requirements.txt

3. Install and start Ollama:

# Download Ollama from https://ollama.com/download
# Then pull the CodeLlama model:
ollama pull codellama:7b

4. Start Ollama server:

ollama serve

python -m app.build_index --github https://github.com/pallets/flask

python -m app.build_index --local /path/to/your/repo

streamlit run ui/streamlit_app.py

Open http://localhost:8501 in your browser.

From GitHub:

python -m app.build_index --github https://github.com/tiangolo/fastapi

From Local Path:

python -m app.build_index --local ./my-project

python test_rag.py

1. Start Streamlit: streamlit run ui/streamlit_app.py
2. Enter your question
3. Click "Ask"
4. View AI-generated answer and source code

┌─────────────┐
│   GitHub    │
│  Repository │
└──────┬──────┘
       │
       ▼
┌─────────────────┐
│ Code Ingestion  │
│  (AST Parser)   │
└──────┬──────────┘
       │
       ▼
┌──────────────────┐
│ Code Chunks      │
│ (Functions/      │
│  Classes)        │
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│ Embeddings       │
│ (SentenceTrans-  │
│  former)         │
└──────┬───────────┘
       │
       ▼
┌──────────────────┐
│ FAISS Index      │
│ (Vector Store)   │
└──────┬───────────┘
       │
       ▼
┌──────────────────┐      ┌─────────────┐
│ User Question    │─────▶│   Search    │
└──────────────────┘      └──────┬──────┘
                                 │
                                 ▼
                          ┌──────────────┐
                          │  Retrieved   │
                          │    Code      │
                          └──────┬───────┘
                                 │
                                 ▼
                          ┌──────────────┐
                          │   Ollama     │
                          │ (CodeLlama)  │
                          └──────┬───────┘
                                 │
                                 ▼
                          ┌──────────────┐
                          │    Answer    │
                          └──────────────┘

codebase-rag/
├── app/
│   ├── __init__.py
│   ├── vector_store.py       # FAISS vector store
│   ├── ingest_code.py         # Code extraction (AST)
│   ├── ingest_github_repo.py  # GitHub cloning
│   ├── build_index.py         # Index building pipeline
│   └── rag_answer.py          # RAG system with Ollama
├── ui/
│   └── streamlit_app.py       # Web interface
├── data/
│   ├── repos/                 # Cloned repositories
│   └── code_index/            # FAISS index + metadata
├── test_rag.py                # CLI test script
├── requirements.txt           # Dependencies
└── README.md                  # This file

Edit app/rag_answer.py:

# Default: codellama:7b
# Other options: codellama:13b, deepseek-coder, starcoder
rag = RAGAnswerer(model="deepseek-coder")

In the Streamlit UI sidebar, use the slider to change the number of code chunks retrieved (default: 5).

Edit app/ingest_code.py:

MAX_FILE_SIZE_KB = 500        # Skip files larger than this
MAX_CHARS_PER_FILE = 100000   # Character limit per file

Run the test script:

python test_rag.py

This will:

1. Load the index
2. Ask 3 test questions
3. Show answers and retrieved code

Solution: Build an index first using python -m app.build_index

Solution: Start Ollama with ollama serve, then pull the model with ollama pull codellama:7b

Solution: The repository might not have any Python files, or they're all being filtered out. Check the ignore lists in app/ingest_code.py

Solution:

• Use a smaller model: codellama:7b instead of codellama:13b
• Reduce retrieval count (k parameter)
• Index fewer files

1. Code Ingestion: Python files are parsed using AST to extract functions and classes
2. Embedding Generation: Each code chunk is converted to a 384-dim vector using sentence-transformers
3. Vector Indexing: Vectors are stored in a FAISS index for fast similarity search
4. Query Processing: User questions are embedded and searched against the index
5. Context Retrieval: Top-k most similar code chunks are retrieved
6. Answer Generation: Retrieved code + question are sent to Ollama (CodeLlama)
7. Response: AI generates a contextual answer based on actual code

• Indexing Speed: ~100 files/minute (depends on file size)
• Search Latency: <100ms for retrieval
• Answer Generation: 5-15 seconds (depends on model and hardware)
• Memory Usage: ~2GB RAM for small repos, ~5GB for large repos

Contributions welcome! Areas for improvement:

• Support for more languages (JavaScript, Java, etc.)
• Better code chunking strategies
• Web-based index building UI
• Multi-repo indexing
• Code similarity visualization

MIT License - feel free to use this for any purpose.

• FAISS by Meta AI
• sentence-transformers by UKPLab
• Ollama for easy local LLM deployment
• Streamlit for the awesome UI framework

Questions? Open an issue on GitHub!

Built with ❤️ for developers who want to understand codebases faster