RAG Application

A simple Retrieval-Augmented Generation (RAG) app built with LangChain, FastAPI, and plain HTML/JS.

Upload a PDF or text document, then ask questions — the app retrieves relevant chunks and uses an LLM to answer.

---

Architecture

┌──────────┐       ┌──────────────┐       ┌────────────┐
│  Browser  │──────▶│   FastAPI     │──────▶│  LangChain │
│  (HTML)   │◀──────│   Backend     │◀──────│  + FAISS   │
└──────────┘       └──────────────┘       └────────────┘
                          │                       │
                     Upload doc              OpenAI API
                     /query                 (embeddings + chat)

How RAG Works (simplified)

1. Upload — The document is split into small overlapping chunks.
2. Embed — Each chunk is converted to a vector using OpenAI Embeddings.
3. Store — Vectors are stored in an in-memory FAISS index.
4. Query — The user's question is embedded, the top-k similar chunks are retrieved, and passed as context to the LLM which generates an answer.

---

Setup

1. Clone / copy the project

RAG/
├── main.py              # FastAPI backend
├── static/
│   └── index.html       # Frontend
├── requirements.txt
├── .env.example
└── README.md

2. Create a virtual environment

cd RAG
python3 -m venv venv
source venv/bin/activate   # On Windows: venv\Scripts\activate

3. Install dependencies

pip install -r requirements.txt

4. Set your OpenAI API key

cp .env.example .env
# Edit .env and paste your real API key

5. Run the app

uvicorn main:app --reload

Open http://127.0.0.1:8000 in your browser.

---

Usage

1. Click Upload and select a .pdf or .txt file.
2. Type a question in the text box and click Ask.
3. The answer and source chunks will appear below.

---

Key Concepts for Students

Concept	Where in code
Document loading	load_document() — uses LangChain's PyPDFLoader / TextLoader
Text chunking	build_vector_store() — RecursiveCharacterTextSplitter
Embeddings	OpenAIEmbeddings() converts text → vectors
Vector store	FAISS.from_documents() — similarity search index
Retrieval chain	RetrievalQA.from_chain_type() — retrieves context + generates answer
API endpoint	FastAPI @app.post("/query")

---

Notes

• This uses in-memory FAISS — data is lost on restart.
• Uses gpt-3.5-turbo by default. Change the model in get_qa_chain().
• For production, add authentication, persistent storage, and rate limiting.