📚 PaperMind — Research Paper Q&A Engine

Ask anything about any research paper. Get answers grounded in the actual content — with source citations.

---

🔗 Live Demo

research-paper--app.streamlit.app

---

🧠 What is PaperMind?

Most AI tools for research papers work at the surface level — summarizing abstracts or searching across databases of millions of papers.

PaperMind goes deeper.

Upload any research paper PDF → it gets chunked, embedded, and stored in a vector database → you ask questions in natural language → the most relevant sections are retrieved → LLaMA 3.1 generates a precise, grounded answer with source citations.

How PaperMind differs from existing tools

Tool	What it does well	The gap
Bohrium AI	Searches 170M+ papers with citation-backed answers	Literature discovery tool — not designed for deep Q&A on your specific uploaded PDF
Elicit AI	Research question synthesis across thousands of papers	Works at abstract/metadata level, not section-level content
Connected Papers	Visualizes citation graphs to find related papers	Zero Q&A capability — discovery only, not comprehension
SciSpace	AI sidebar that explains highlighted text	Closed platform — no control over LLM, embeddings, or retrieval pipeline
PaperMind	Deep Q&A on your specific uploaded paper	Full transparency — see exactly which chunks were retrieved and why

---

✨ Features

• PDF Upload — drag and drop any research paper, text extracted automatically
• RAG Pipeline — retrieval-augmented generation, not full-document prompting
• Semantic Search Tab — search paper chunks by meaning, not just keywords
• Auto Paper Summary — problem, approach, key findings, keywords, difficulty level
• Source Citations — every answer shows which chunks it retrieved with relevance scores
• RAG Stats Dashboard — architecture visualization, vector counts, query metrics
• No Token Limits — only the top-5 relevant chunks are sent to the LLM, not the full paper
• Proper Section for Images - Integrated figure extraction and caption matching enabling multimodal document understanding.

---

🏗 RAG Architecture

PDF Upload
    ↓
PyMuPDF — text extraction
    ↓
Chunking — 500 words, 100-word overlap
    ↓
sentence-transformers — all-MiniLM-L6-v2 (384-dim embeddings)
    ↓
Pinecone — cosine similarity vector index
    ↓
──────────────────────── Query Time ────────────────────────
User Question → embed → Pinecone top-5 search
    ↓
Retrieved chunks as context
    ↓
LLaMA 3.1 8B via Groq → grounded answer with citations

Why RAG instead of sending the full paper?

Approach	Token usage	Accuracy	Works on long papers
Full document in prompt	Very high	Lower (lost in middle)	❌ Often exceeds limits
RAG (PaperMind)	Low (5 chunks only)	Higher (focused context)	✅ Any length

---

🛠 Tech Stack

Component	Technology
LLM	LLaMA 3.1 8B via Groq API
Embeddings	sentence-transformers all-MiniLM-L6-v2 (free, local)
Vector Database	Pinecone (serverless, cosine similarity)
PDF Parsing	PyMuPDF (fitz)
UI	Streamlit
Environment	Python 3.9+

---

⚡ Run Locally

1. Clone the repo

git clone https://github.com/anubhav9369/research-paper-qa.git
cd research-paper-qa

2. Create virtual environment

python3 -m venv venv
source venv/bin/activate  # Mac/Linux
# venv\Scripts\activate   # Windows

3. Install dependencies

pip install -r requirements.txt

4. Set up API keys

cp .env.example .env

Edit .env and fill in your keys:

PINECONE_API_KEY=your_pinecone_api_key_here
GROQ_API_KEY=your_groq_api_key_here
GEMINI_API_KEY=your_gemini_api_key_here  # optional, not used currently

5. Run

python3 -m streamlit run app.py

Open http://localhost:8501

---

🔑 Getting API Keys (all free)

Key	Where to get it
PINECONE_API_KEY	pinecone.io → Sign up → API Keys
GROQ_API_KEY	console.groq.com → Create API Key

Note: sentence-transformers runs locally — no API key needed for embeddings.

---

📁 Project Structure

research-paper-qa/
├── app.py                  # Streamlit UI — landing page + dashboard
├── src/
│   ├── __init__.py
│   ├── rag_pipeline.py     # Embeddings + Pinecone upsert/search
│   ├── pdf_parser.py       # PDF text extraction + section detection
│   └── llm_qa.py           # RAG-grounded Q&A + paper summary generation
├── requirements.txt
├── .env.example
├── .gitignore
└── README.md

---

🖥 App Screenshots

Landing Page

Upload-first design — drop your PDF and start asking questions immediately.

Paper Dashboard

After processing, you get an auto-generated overview:

• One-liner description
• Problem and approach
• Key findings
• Keywords
• Difficulty level (Beginner / Intermediate / Advanced)
• Field of research

Ask Questions Tab

Natural language Q&A with source chunk citations and relevance scores.

Semantic Search Tab

Search paper chunks directly by semantic similarity — useful for finding specific sections.

RAG Stats Tab

Full architecture diagram + vector counts + session query metrics.

📄 License

MIT License — free to use, modify, and build on.