An intelligent Retrieval-Augmented Generation (RAG) system that enables developers to ask natural language questions about any GitHub repository and receive context-aware answers grounded in the repository's source code.

Repository RAG Assistant combines:

• AST-based code chunking
• Semantic embeddings
• FAISS vector search
• Cross-Encoder reranking
• Large Language Models (Qwen 2.5 Coder)

to create an AI-powered code understanding system capable of answering repository-specific questions.

Instead of manually searching through hundreds of files, developers can ask:

• How does OAuth authentication work?
• What does make_not_authenticated_error() do?
• How are security scopes handled?
• Which files are involved in user authentication?

and receive answers grounded in the repository source code.

• AST-based Python parsing
• Function-level chunking
• Method-level chunking
• Class metadata extraction
• Duplicate reduction

• BAAI BGE embeddings
• FAISS vector indexing
• Similarity search
• Fast retrieval

• Cross-Encoder reranking
• Improved retrieval quality
• Better repository grounding

• Qwen2.5-Coder integration
• Context-aware code explanations
• Repository-focused responses
• Source-backed answers

• FastAPI backend
• Swagger documentation
• JSON request/response format

GitHub Repository
        │
        ▼
AST Code Chunking
        │
        ▼
Code Chunks
        │
        ▼
BGE Embeddings
        │
        ▼
FAISS Vector Store
        │
        ▼
Retriever
        │
        ▼
Cross Encoder Reranker
        │
        ▼
Qwen2.5-Coder
        │
        ▼
Generated Answer

• Python
• FastAPI
• Pydantic

• FAISS
• Sentence Transformers

• BAAI/bge-small-en-v1.5

• cross-encoder/ms-marco-MiniLM-L-6-v2

• Qwen/Qwen2.5-Coder-3B-Instruct

backend/
│
├── api/
│   ├── routes.py
│   └── schemas.py
│
├── embeddings/
│   ├── embedder.py
│   └── build_index.py
│
├── ingestion/
│   ├── chunker.py
│   └── build_chunks.py
│
├── llm/
│   ├── model_loader.py
│   ├── generator.py
│   └── prompt.py
│
├── rag/
│   └── pipeline.py
│
├── retrieval/
│   ├── retriever.py
│   └── reranker.py
│
└── main.py

tests/
data/
docs/

git clone <repository-url>
cd repository-rag-assistant

python -m venv venv

Windows:

venv\Scripts\activate

Linux/Mac:

source venv/bin/activate

pip install -r requirements.txt

python -m backend.ingestion.build_chunks

python -m backend.embeddings.build_index

uvicorn backend.main:app

Server:

http://127.0.0.1:8000

Swagger UI:

http://127.0.0.1:8000/docs

Request:

{
  "question": "What does make_not_authenticated_error do?"
}

Response:

{
  "answer": "The make_not_authenticated_error method creates an HTTPException with status code 401 and sets the WWW-Authenticate header..."
}

• How does OAuth authentication work?
• What is the purpose of Security()?
• How are OAuth scopes handled?
• Explain the authentication flow.
• Which files implement API key security?
• How does dependency injection work?

Current pipeline:

• AST-based chunking
• 588 indexed chunks
• FAISS vector retrieval
• Cross-Encoder reranking
• Qwen 2.5 Coder generation

• Multi-language repository support
• Multi-repository indexing
• Hybrid Search (BM25 + Vector Search)
• Docker deployment
• React frontend
• Conversation memory
• GitHub repository ingestion via URL
• Source citations in responses
• GitHub Action integration

MIT License