Paper-2X

Paper2X is an AI-powered tool that leverages Gemini 1.5 Flash and LangGraph, to transform research papers into engaging presentations and audio summaries. The project provides a backend service to process document inputs and convert them into structured PPTs.

Project Motivation

Understanding and presenting research papers can be time-consuming. Paper2X streamlines this process by automating content extraction, slide creation, and audio generation, making research more accessible to students, educators, and professionals.

Also, this project is a continuation from my team's partial solution for the Research Remix track sponserd by Cactus Communications during the MineD Hackathon 2025, where some features of this project were the expected outcomes.

Table of Contents

• Overview
• Features
• Architecture
• Installation
• Usage
  • Running the API
  • Frontend Integration
• Configuration
• Dependencies
• Contributing
• License
• Contact

Overview

The goal of Paper-2-X is to streamline the creation of presentation slides from academic papers or long-form documents. By combining the power of FastAPI with state-of-the-art language models (via LangChain and its ecosystem), the application parses input documents, extracts key points, and generates a coherent slide deck.

Features

• Research Paper to Presentation: Automatically generates PowerPoint presentations (in .pptx) from academic papers. Uses:
  • PyMuPDFLoader to extract text and images from the research articles
  • Gemini 1.5 Flash to generate slides' content for the PPT
  • pptx library for creating the PPT from text. Includes support for a few themes, which are created using pptx library itself.
• Research Paper to Podcast Transcript: Converts the research paper content into a transcripts for a narrated-like podcast.
• API Deployment: This entire LangGraph workflow is deployed as a FastAPI app, and has endpoints to get the specific content (check Running The API section)

Architecture

The application follows a modular design:

• API Layer: Implemented with FastAPI and served using Uvicorn.
• Processing Agents: Under the agents directory, different agents handle subtasks such as text extraction, summarization, and PPT slide creation.
• Configuration: Environment variables (using python-dotenv) allow configuration for API keys and other secrets.

Installation

1. Clone the repository:

   git clone https://github.com/lavanderhoney/paper2x.git
   cd paper2x

2. Create and activate a virtual environment (optional but recommended):

   python -m venv venv
   source venv/bin/activate   # On Windows use: venv\Scripts\activate

3. Install the required packages:

   pip install -r requirements.txt

Usage

This project contains an API which serves the langgraph workflow, but not the PPT generation and the podcast. The API is experimental only.

To use the features, run the agents\workflow_testing.ipynb notebook.

Running the API

uvicorn app.main:app --reload

Access the API at http://127.0.0.1:8000/docs

Endpoints

You can test these endpoints using Postman (as there's no front-end, yet).

• POST /generate_ppt
  • Upload the research paper as a form data in the body of POST request in the Postman API interface.
  • This will execute the langgraph workflow, and return a file_id. Save this id, as it will be required for further GET requests.
• GET/ppt/{file_id}
  • Hitting this endpoint, using the file_id obtained from the POST request, will return the textual contents for creating the PPT
• GET/summary/{file_id}
  • Returns the summary of the research paper, in a form useful for generating the podcast transcript.
• GET/convo/{file_id}
  • Returns the transcript of the podcast, in a JSON containing lists of dialagoues of two personas: Katherine, and Clay, and a list specifying the order of their dialogues. The datamodel is specified in the agents\agent.py as the Conversation class.

NOTE: The purpose behind using the file_id thing, is to have a way to store the graph's state in-memory of the API. This violates the property of RESTful API being stateless, but I've implemented this as a temporary way to circumvent executing the graph mutliple times by the same user.

File Structure

📂 paper2x
├── agents
│   ├── extracted_images # Stores the images extracted from the pdf
│   ├── static # For storing the generated ppts
│   ├── uploaded_pds # Research papers to be used
│   ├── agent.py        # Python script for the agentic workflow
│   ├── app2.py  # FastAPI app entry point
│   └── workflow_testing.ipynb # Jupyter notebook for agentic workflow testing
├── requirements.txt
└── README.md

Configuration

Create a .env file and add the following keys:

GOOGLE_API_KEY=your_gemini_api_key
LANGSMITH_API_KEY=your_langsmith_api_key
LANGSMITH_ENDPOINT=your_langsmith_endpoint
LANGSMITH_PROJECT=paper2x_project
LANGSMITH_TRACING=true

Dependencies

The project depends on a number of Python packages:

• Backend & API:
  • fastapi
  • uvicorn
• AI Agents:
  • langchain
  • langgraph
  • langchain_google_genai
  • langchain_core
• Utilities:
  • pydantic
  • python-dotenv
  • python-multipart
  • PyMuPDF
  • typing-extensions
  • uuid

See requirements.txt for the complete list.

Contributing

Contributions are welcome! If you’d like to help improve Paper-2-X:

1. Fork the repository.
2. Create a new branch for your feature or bugfix.
3. Commit your changes.
4. Open a pull request detailing your changes.

Please follow the existing code style and include tests for new functionality.

License

This project is licensed under the MIT License.

Contact

For questions, suggestions, or contributions, please open an issue on GitHub or contact the repository owner at your preferred communication channel.