discuss

dicuss is an interactive discussion application powered by Large Language Models (LLMs). It enables you to simulate multi-entity debates on any topic, where each entity can represent a unique persona, expert, or fictional character. Entities draw their knowledge from uploaded PDF documents or Wikipedia articles, and participate in turn-based discussions, responding to each other's arguments in lively, multi-cycle debates.

Features

• Simulated Multi-Agent Debates: Entities (LLM-powered) discuss user-provided topics over multiple rounds, each adapting their arguments as the conversation evolves.
• Persona Mode: Assign Wikipedia pages to entities, and they will debate topics as if they are that historical figure, expert, or concept—complete with their own knowledge and style.
• Custom Knowledge Bases: Upload PDFs or link Wikipedia articles to provide entities with unique sources of information.
• Interactive Web UI: Built with Streamlit for a seamless, visual, and real-time debate experience.
• Dynamic Contextual Arguments: Each entity references previous statements and loaded documents, making the debate context-aware and engaging.
• Easy Material Management: Load and track PDF/Wiki materials for all entities via a sidebar interface.
• Containerized Deployment: Includes a Dockerfile for quick setup and reproducibility.

Getting Started

Prerequisites

• Python 3.11+
• (Optional) Docker

Installation

1. Clone the repository:

   git clone https://github.com/yarlaw/discuss
   cd discuss

2. Install dependencies:

   pip install -r requirements.txt
   pip install streamlit

3. Run the application:

   streamlit run streamlit_app.py

   The app will be available at http://localhost:8501.

Docker

1. Build the container:

   docker build -t stream-chat .

2. Run the container:

   docker run -p 8501:8501 stream-chat

Usage

1. Add Entities: Each entity can be a generic AI, or be assigned a specific Wikipedia page and/or PDF sources.
2. Configure Materials: Upload PDFs or add Wikipedia links for each entity. Materials are loaded and indexed for contextual use.
3. Set Topic & Cycles: Choose a discussion topic and how many rounds ("cycles") of debate you'd like.
4. Start the Debate: Watch as entities take turns, referencing their sources and each other's arguments.
5. Persona Mode: Entities with Wikipedia links will argue as that persona, using knowledge and style from their Wiki article.

Example Scenarios

• Historical Debates: Let Einstein, Tesla, and Marie Curie debate the future of technology.
• Policy Arguments: Simulate expert panels with uploaded reports as their knowledge base.
• Fictional Roundtables: Have Gandalf, Sherlock Holmes, and Yoda debate the meaning of wisdom.

Configuration

• Materials Folder: Uploaded PDFs are stored in RAG_files/.
• Model Selection: Default LLM is mistral-7b (configurable in code).
• Discussion Cycles: Adjustable via the UI sidebar.

Technologies Used

• Streamlit
• LangChain
• HuggingFace Transformers
• FAISS (for document indexing)
• [PyPDF2, PyMuPDF, BeautifulSoup4] (for parsing documents)