📄 Retrieval-Augmented Chatbot API

A FastAPI-based RAG (Retrieval-Augmented Generation) chatbot that answers user queries based on information extracted from a local resume.json.
Supports rate-limiting, CORS for frontend integration, and retrieval with Hugging Face embeddings + Chroma vector store.

---

🚀 Features

✅ Retrieval-Augmented Generation using langgraph
✅ OpenAI Sentiment embeddings ✅ Chroma vector database for document retrieval
✅ FastAPI REST API ready for frontend integration
✅ CORS configured for specific frontend URLs
✅ Rate-limited endpoint to prevent abuse (10 requests per hour per IP)
✅ Cleaned AI responses (removes <think> tags)

---

📂 Project Structure

.
├── main.py                 # FastAPI app and chatbot API
├── resume.json             # JSON knowledge base for retrieval
├── requirements.txt
└── README.md

---

⚙️ Technologies Used

• FastAPI - RESTful API framework
• slowapi - Rate limiting
• langchain_huggingface / langchain_chroma - Embeddings & Vector DB
• Groq Deepseek model - LLM backend
• RecursiveCharacterTextSplitter - Chunking JSON documents
• LangChain Hub Prompt - For RAG prompt template
• CORS Middleware - Frontend integration

---

🔑 API Endpoint

POST /ask

• Description: Ask the chatbot a question and get a generated answer
• Rate Limit: 10 requests per hour per IP
• Request JSON:

{
  "message": "Your question here"
}

• Response JSON:

{
  "answer": "AI generated answer based on resume knowledge"
}

---

🌐 CORS Allowed Origins:

• http://localhost:5173
• https://pandalow.github.io/

---

📖 How It Works (RAG Pipeline)

1. Load and split resume.json
2. Store chunks in a Chroma vector database
3. Accept user question
4. Perform similarity search to retrieve relevant context
5. Format context and question into a RAG prompt
6. Generate response using the LLM (deepseek-r1-distill-qwen-32b)
7. Clean response (remove <think> tags)
8. Return the answer

---

🛠 Running the Project

1️⃣ Install Requirements

pip install -r requirements.txt

2️⃣ Start the API Server

uvicorn main:app --reload

---

📈 Example Usage (curl)

curl -X POST "http://localhost:8000/ask" \
     -H "Content-Type: application/json" \
     -d '{"message": "Tell me about your backend skills."}'

---

⚠️ Rate Limiting

• Limit: 10/hour per IP
• Returns 429 Too Many Requests if exceeded

---

📜 Notes

✅ LLM model used: deepseek-r1-distill-qwen-32b via Groq
✅ Embedding model: sentence-transformers/paraphrase-MiniLM-L6-v2
✅ Vector DB: Chroma
✅ JSON schema customizable via JQ syntax

---

🤝 License

MIT License — Feel free to use and modify.