# Self-RAG Agent

A Retrieval-Augmented Generation system that grades its own retrieved

documents and generated answers before responding — reducing hallucination

and irrelevant-context answers compared to standard RAG.

## How it works

1. **Retrieve** — pull top-k chunks from a FAISS vector store built from your PDF/notes

2. **Grade documents** — LLM checks each chunk for relevance to the question, discards irrelevant ones

3. **Generate** — answer is produced using only the relevant chunks

4. **Grade hallucination** — checks the answer is grounded in the retrieved context

5. **Grade usefulness** — checks the answer actually addresses the question

6. Retries generation (up to 2 times) if either check fails; returns "I don't know" if no relevant documents are found

## Setup

python -m venv venv

venv\\Scripts\\activate

pip install -r requirements.txt

cp .env.example .env   # add your GROQ\_API\_KEY

Put a PDF or .txt file in data/sample.pdf (or update the path in the test scripts).

## Run

python tests/test\_rag.py        # basic retriever test

python tests/test\_self\_rag.py   # full self-correcting RAG loop

## Stack

LangChain, Groq (Llama 3.3 70B), FAISS, HuggingFace sentence-transformers, Pydantic