Pegasus

Pegasus is an AI-driven financial analysis platform designed to streamline document ingestion and reporting. It leverages large language models (LLMs) for both text and multimodal (vision) tasks, enabling complex PDF ingestion, chunk-wise summaries, and robust interactive chat features.

---

Table of Contents

• Background
• Features
• Tech Stack
• Installation
  • Prerequisites
  • Server Setup (Python)
  • UI Setup (Nextjs)
  • Running the Application
• Usage
• Contribution Guide
• Future Improvements
• License

---

Background

Financial analysts frequently work with large, unstructured documents (PDFs, Word documents, etc.) and spend significant time extracting key metrics and commentary to produce financial reports. Pegasus automates and augments this process by:

1. Indexing documents (supporting PDF, DOC, DOCX) for quick retrieval. This uses the ColPali pipeline for all kinds of documents.
2. Generating chunk-based summaries for multi-page files using a vision-language model (Qwen2-VL).
3. Creating a final, multi-section financial analysis report with minimal user input.
4. Offering a user-friendly chat interface to query the ingested documents and retrieve relevant context.

---

Features

• Document Ingestion & Indexing
  Upload PDF/DOC/DOCX documents, automatically convert them to text or PDF, and index them using Byaldi’s RAGMultiModalModel for retrieval.

• RAG (Retrieval-Augmented Generation)
  Quickly fetch relevant pages/sections from ingested PDFs.

• Vision + Text Models
  Summaries are generated by a large vision-language model, while text-only LLMs (Qwen2.5 series) create the final cohesive report.

• Interactive Chat
  Query the AI about financial or general questions; the system automatically retrieves supporting context from the indexed documents.

• Report Generation
  Generate a structured financial analysis report (executive summary, key metrics, detailed analysis, trends, and conclusions).

---

Tech Stack

Server

• Language & Framework: Python 3.11 + FastAPI
• Core Libraries:
  • Byaldi for RAG indexing
  • Transformers for model loading
  • docx2pdf for document conversion
  • PyMuPDF (fitz) for PDF processing

UI

• Language & Framework: TypeScript, Next.js 15 (App Router)
• Styling: Tailwind CSS
• UI Components: Radix UI, ShadCN/UI libraries
• State Management: React hooks & local state
• Other Libraries:
  • React Dropzone for file uploads
  • React Markdown / Showdown for Markdown rendering
  • Plotly.js / Chart.js for prospective data visualization

Installation

Prerequisites

1. Python 3.11 and uv (the code specifically targets 3.11)
2. Node.js & npm (latest LTS recommended, e.g., Node 18+)
3. Git

---

Server Setup (Python)

1. Clone the Repository:

git clone https://github.com/Arrabonae/pegasus.git
cd pegasus

2. Create & Activate a Virtual Environment (optional but recommended):

uv venv
source venv/bin/activate  # On Mac/Linux

3. Install Python Dependencies:

uv pip install -r requirements.txt

4. Run the Server:

uvicorn server.main:app --host 0.0.0.0 --port 5050 --reload

The server will be available at http://localhost:5050.

UI Setup (Next.js)

1. Install Node Dependencies:

npm install

2. Run the Development Server:

npm run dev

The UI will be running on http://localhost:3000.

Running the Application

1. Start the Python (FastAPI) Server on port 5050.
2. Start the Next.js Dev Server on port 3000.
3. In your browser, navigate to http://localhost:3000. The UI communicates with the backend at http://localhost:5050.

Usage

1. Create a New Thread
   • Click “New Thread” in the sidebar.
   • Upload a PDF/DOC/DOCX file, provide a descriptive title, and start indexing.
2. Add Additional Files
   • In the right-hand “Files” tab, use the paperclip icon to upload more files to the current session.
3. Chat with the AI
   • Use the chat box at the bottom-right corner to ask questions.
   • The AI retrieves relevant pages from your indexed documents, providing answers with references.
4. Generate a Report
   • In the main “Report” panel, click “Generate Report.”
   • A multi-section report (executive summary, key metrics, detailed analysis, trends, and conclusions) is compiled by the vision + text LLM pipeline.

Contribution Guide

We welcome your contributions and feedback! Here’s how you can get involved:

1. Fork the repository and clone to your local machine.
2. Create a new branch for your feature or bug fix:

git checkout -b feature/your-feature

3. Commit your changes with clear messages:

git commit -m "Add awesome feature"

4. Push to your fork and create a Pull Request from GitHub.

We encourage discussions around project architecture, code style, or new features. Feel free to open an issue or start a GitHub Discussion.

Future Improvements

Below are some prioritized enhancements the team is looking to implement:

1. Implementation of Charts
   • Integrate dynamic charts for financial metrics using Plotly.js or Chart.js, enabling real-time visual analytics in the UI.
2. Implementation of Experimental Features
   • Enable advanced or less-tested functionality behind feature flags.
   • Expand the “Experimental” toggle to unlock new UI interactions or server endpoints.
3. Implementation of Yahoo Finance Features
   • Integrate external financial data from Yahoo Finance, merging real-time market data with user-provided documents.
4. Implementation of User-Directed Report Changes
   • Allow users to manually amend or annotate the AI-generated final report.
   • Provide an interface for “edit suggestions” that automatically merges user edits into the final text.

License

This project is licensed under the APACHE LICENSE, VERSION 2.0.