b05un/school_rag_api

📚 FUOYE RAG API — Hugging Face Space Deployment

This project is a Retrieval-Augmented Generation (RAG) API service for Federal University Oye-Ekiti (FUOYE), built with FastAPI, LangChain, OpenAI, Chroma, and containerized with Docker. It’s deployed on Hugging Face Spaces.

The app answers questions using pre-loaded university documents, providing helpful, friendly, and relevant responses — or redirecting users politely when information is unavailable.

📦 Project Structure

.
├── app.py               # Main FastAPI app and RAG logic
├── Dockerfile           # Docker configuration for deployment on HF Spaces
├── data/                # Directory containing .txt documents to index
├── .env                 # Environment variables (OPENAI_API_KEY etc.)
└── chroma_db/           # Local Chroma vectorstore directory (auto-generated)

⚙️ How It Works

📌 Environment Setup

• —The app loads environment variables using python-dotenv.
• —Important variables:
• —DATA_PATH — path to directory containing text documents
• —CHROMA_PATH — directory for the Chroma vector database
• —OPENAI_API_KEY — your OpenAI API key for embeddings and LLM

Defaults for Hugging Face Spaces:

• —DATA_PATH=/code/data
• —CHROMA_PATH=/tmp/chroma_db

📌 Document Loading & Splitting

1. 1.DirectoryLoader loads all .txt files from the data/ directory.
2. 2.RecursiveCharacterTextSplitter splits documents into chunks of 1000 characters with 150 characters overlap — improving retrieval granularity.

loader = DirectoryLoader(DATA_PATH, glob="**/*.txt", loader_cls=TextLoader)
splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150)

📌 Embedding & Vector Store Creation

• —Uses OpenAI Embedding model `text-embedding-3-large` to generate embeddings.
• —Chunks are stored in a Chroma vectorstore locally at CHROMA_PATH.
• —Vectorstore is created on app start via get_vector_store().

📌 API Endpoints

Base URL: /

📡 /rag-query Endpoint Workflow

Request Body:

{
  "message": "Your question here",
  "history": "Previous conversation if any"
}

Response:

{
  "response": "RAG-based answer"
}

Process:

1. 1.Retrieve Top-5 Relevant Documents using Chroma retriever.
2. 2.Construct RAG Prompt — combines:
3. 3.System instructions for a helpful FUOYE assistant
4. 4.User’s question
5. 5.Conversation history
6. 6.Retrieved knowledge content
7. 7.LLM Streaming Response using OpenAI gpt-4o-mini
8. 8.Return Final Answer as JSON.

📦 Docker Deployment (on Hugging Face Spaces)

Dockerfile Highlights:

• —Python 3.11 base image.
• —Installs dependencies from requirements.txt
• —Exposes port 7860 for Hugging Face compatibility.
• —Runs the app using uvicorn.

To Deploy:

1. 1.Push code to Hugging Face repository.
2. 2.Set HF API Token and OPENAI_API_KEY in repository secrets.
3. 3.Hugging Face automatically builds Docker image and runs container.

📎 Example .env

OPENAI_API_KEY=sk-yourkeyhere
DATA_PATH=/code/data
CHROMA_PATH=/tmp/chroma_db

✅ Usage Notes

• —Update data/ with .txt files containing FUOYE knowledge.
• —On app start, the Chroma vectorstore rebuilds automatically.
• —Adjust chunk size and retrieval k as needed for performance tuning.

✨ Technologies Used

• —FastAPI
• —LangChain
• —OpenAI API
• —Chroma Vector Database
• —Docker
• —Hugging Face Spaces

📖 License

MIT License — free for educational, research, and personal projects.