DocMind an AI-Powered PDF Q&A app built with RAG(retrieval-Augmented Generation)

• LIVE DEMO : https://docmind-ragmind.streamlit.app/

what id DocMind? DocMind lets you upload any PDF and have a conversation with it. ask questions, get accurate answers with page citations, and summarize entire documents in sseconds-all powered by a full RAG pipeline built from scratch.

features: *pdf upload *semantic search-finds relevent passsages using meaning , not just keywords *Ai answers - poweres by Groq's llama-3.1-8b *source citations-eevry answer includes page numbers and source chuncks *document summarization by one click. *handles spelling mistakes -pular/singular. *chat interface

HOW IT WORKS:

PDF Upload ↓ Text Extraction (PyMuPDF) ↓ Sentence-Aware Chunking (500 chars, overlap) ↓ Embeddings (all-MiniLM-L6-v2, 384 dimensions) ↓ Vector Store (cosine similarity search) ↓ Query → Embed → Search → Top Chunks ↓ Groq llama-3.1-8b → Answer with citations

Tech Stack Component Technology Frontend Streamlit PDF Parsing PyMuPDF (fitz) Embeddings sentence-transformers (all-MiniLM-L6-v2) Vector Store Custom cosine similarity (numpy) LLM Groq API — llama-3.1-8b-instant Deployment Streamlit Cloud

Project Structure:

rag-doc-qa/ ├── app.py # Streamlit UI ├── pdf_parser.py # PDF text extraction + cleaning ├── chunker.py # Sentence-aware chunking with overlap ├── embedder.py # SentenceTransformer embeddings ├── vector_store.py # Custom vector store with cosine search ├── llm.py # Groq LLM integration + prompt engineering ├── requirements.txt # Dependencies └── README.md

Clone the repo

git clone https://github.com/ankitha-km/rag-doc-qa.git cd rag-doc-qa

Create virtual environment

python -m venv venv venv\Scripts\activate # Windows

source venv/bin/activate # Mac/Linux

Install dependencies

pip install -r requirements.txt

Set up environment variables

Create a .env file with:

GROQ_API_KEY=your_groq_api_key_here

Run the app

streamlit run app.py

Get a free Groq API key at console.groq.com

Key Learnings Building this project covered the full RAG pipeline from scratch:

Why RAG? LLMs have context limits and no access to private documents. RAG solves both. Embeddings convert text meaning into vectors — similar meaning = similar vectors Chunking tradeoff — too small loses context, too large loses precision Cosine similarity measures angle between vectors to find semantically similar chunks Prompt engineering — strict vs flexible prompts dramatically affect answer quality LLM choice matters — 0.5B models hallucinate, 8B models reason properly

Future Improvements

Multi-PDF support Chat memory (conversation context) Persistent vector store across sessions Streaming responses Support for DOCX, TXT files

Author Built by Ankitha as a learning project to understand RAG systems from the ground up.