PaperLens

AI-powered research paper analysis and Q&A engine.

Upload up to 3 research papers (PDF) and get structured insights extracted by an LLM, plus a conversational interface to ask questions across your uploaded corpus.

---

Features

• Batch upload — up to 3 PDFs per session
• Structured extraction — title, authors, abstract, problem statement, methods, math formulas (LaTeX), datasets, results, limitations, future work, keywords
• Math rendering — formulas rendered via KaTeX in the browser
• RAG Q&A — ask questions across all uploaded papers, grounded in retrieved context
• Conversation memory — last 6 turns kept in context per session
• Session isolation — each upload batch gets its own vector store collection

---

Stack

Layer	Technology
API	FastAPI + Uvicorn
PDF extraction	Docling (OCR disabled)
LLM	Groq (llama-3.3-70b-versatile)
Embeddings	sentence-transformers/all-MiniLM-L6-v2
Vector store	ChromaDB (embedded, persistent)
Frontend	Vanilla HTML/CSS/JS + KaTeX

---

Setup

1. Clone and create environment

git clone <your-repo-url>
cd PaperLens
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate

2. Install dependencies

pip install -r requirements.txt

Note: Docling downloads its ML models (~500MB) on first run. This is a one-time cost.

3. Configure environment

cp .env.example .env
# Edit .env and add your GROQ_API_KEY

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

4. Run

uvicorn app.main:app --reload

Open http://localhost:8000

---

Running Tests

pip install pytest pytest-asyncio
pytest tests/ -v

Tests mock all external services (Docling, Groq, ChromaDB). No API key needed for tests.

To test PDF parsing, place a sample PDF at tests/fixtures/sample.pdf.

---

Project Structure

PaperLens/
├── app/
│   ├── main.py                  # FastAPI app
│   ├── schemas.py               # HTTP-level Pydantic models
│   └── routes/
│       ├── analysis_routes.py   # POST /analyze
│       └── chat_routes.py       # POST /chat
│
├── paper_analyzer/
│   ├── ingestion/
│   │   ├── pdf_parser.py        # Docling extraction
│   │   └── chunker.py           # Section-aware chunking
│   ├── extraction/
│   │   ├── insight_extractor.py # Groq LLM extraction
│   │   └── schemas.py           # PaperInsight model
│   ├── vectorstore/
│   │   ├── embedder.py          # MiniLM embeddings
│   │   └── store.py             # ChromaDB operations
│   └── chat/
│       ├── context_builder.py   # RAG prompt assembly
│       └── chat_llm.py          # Groq chat call
│   └── pipeline.py              # asyncio.gather orchestration
│
├── frontend/
│   └── index.html               # Full UI
│
└── tests/
    ├── test_pdf_parser.py
    ├── test_insight_extractor.py
    ├── test_pipeline.py
    └── test_chat.py

---

API Endpoints

POST /analyze

Upload PDFs and get structured insights.

Request: multipart/form-data with files (1–3 PDF files)

Response:

{
  "session_id": "uuid",
  "papers": [
    {
      "filename": "paper.pdf",
      "title": "...",
      "authors": "...",
      "abstract": "...",
      "problem_statement": "...",
      "methods": "...",
      "math_formulas": [{"formula": "...", "meaning": "..."}],
      "datasets": "...",
      "results": "...",
      "limitations": "...",
      "future_work": "...",
      "keywords": ["..."],
      "year": "2024"
    }
  ]
}

POST /chat

Ask a question about the uploaded papers.

Request:

{
  "session_id": "uuid",
  "question": "What methods are common across the papers?",
  "history": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]
}

Response:

{
  "answer": "...",
  "sources": ["Paper Title A", "Paper Title B"]
}

---

Design Notes

• Parallel processing: All 3 papers are processed concurrently via asyncio.gather(). Total latency ≈ slowest single paper, not the sum.
• Dual parallelism per paper: Groq extraction and ChromaDB population run simultaneously for each paper.
• 3-paper limit: Enforced at the route level with a clear error message.
• "Info not found": Every extraction field defaults to this string so the UI always has consistent data to render.
• Docling OCR disabled: Assumes digitally typeset PDFs (standard for arXiv). Enable do_ocr=True in pdf_parser.py for scanned documents (slower).