convez.ai
AgenteProdutividadeIntermediário
rayansyed95

AI Panel Discussion — Debate Multiagente

rayansyed95/AI-panel-discussion-project

Painel de debate com 3 IA, busca web e memória de contexto inteligente

Ver no GitHub
InicialPluginScore · de 0 a 100

Aplicação de debate entre três modelos de IA (GPT-4o, Gemini e Claude) que discutem tópicos com busca web ao vivo e memória de contexto compartilhada. Ideal para explorar perspectivas diferentes sobre um mesmo tema com moderação humana em tempo real.

Instalação

$npx convez-ai rayansyed95/AI-panel-discussion-project

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
v1.0.0
Issues abertas
1
Contribuidores
1

Casos de uso

  • Gerar múltiplas perspectivas sobre um tópico complexo
  • Pesquisa qualitativa com dados atualizados via web search
  • Criação de conteúdo através de debate estruturado entre modelos

Veredito

Implementação criativa de multi-agent orchestration com Streamlit e APIs públicas. Conceito sólido de memória rolling para reduzir token usage, mas o projeto carece de testes automatizados e está limitado a um arquivo único (app.py). Útil para prototipagem e exploração de dinâmicas entre modelos.

README

A tradução para o português deste README ainda está na fila.

🎙️ AI Panel Discussion

A multi-agent panel debate app built with Streamlit, where GPT-4o, Gemini, and Claude discuss any topic you give them — with live web search, rolling context memory, and human-in-the-loop moderation.

Python Streamlit License


✨ Features

  • Three AI Panelists — each with a distinct persona:
    • 🟢 GPT-4o — concise, logical, data-driven
    • 🔵 Gemini Pro — creative, visionary, big-picture
    • 🟠 Claude 3.5 — nuanced, ethical, devil's advocate
  • Live Web Search via DuckDuckGo — agents can pull real-time context before responding
  • Rolling Summary Memory — a lightweight "crux" of the discussion is maintained and passed to each agent so context stays coherent without ballooning token usage
  • Human-in-the-Loop — inject your own questions or points as Moderator at any time
  • Turn-based Flow — click "Next Turn" to step through agents one at a time, or let the discussion flow naturally

🚀 Getting Started

1. Clone the repo

git clone https://github.com/your-username/ai-panel-discussion.git
cd ai-panel-discussion

2. Install dependencies

pip install streamlit openai anthropic google-genai duckduckgo-search

Note: duckduckgo-search is optional. The app runs without it — web search will be disabled and agents will display a warning instead of search results.

3. Run the app

streamlit run app.py

4. Add your API keys

In the sidebar, expand 🔑 API Keys and paste in the keys for whichever providers you want to use. You can run the app with just one or two keys — panelists without a key will fall back to a mock response.

ProviderWhere to get a key
OpenAIhttps://platform.openai.com/api-keys
Google Geminihttps://aistudio.google.com/app/apikey
Anthropichttps://console.anthropic.com/settings/keys

🖥️ How It Works

User sets a topic
       │
       ▼
Moderator message added to discussion
       │
       ▼
User clicks "Next Turn"
       │
       ├── Agent performs DuckDuckGo web search on the topic
       ├── Agent receives: topic + rolling summary + last message + search results
       ├── Agent calls its LLM API and generates a ≤150-word response
       └── Response is summarised and appended to the rolling context
       │
       ▼
Next agent's turn (cycles: GPT-4o → Gemini → Claude → ...)

The rolling summary (visible in the sidebar) is the key to keeping the discussion coherent. Rather than passing the full transcript to every agent (which would get expensive quickly), each new response is summarised to its core point and appended to a running crux that all agents share.


📁 Project Structure

ai-panel-discussion/
├── app.py          # Main Streamlit application
└── README.md

⚙️ Configuration

SettingWhereDefault
Discussion topicMain UI form"Effect of climate change on rain cycle in Karachi"
API keysSidebar expanderNone (mock mode)
Web search results per queryperform_web_search() — max_results param2
Max response lengthAgent system prompt150 words
Summary truncationsummarize_text()First 3 sentences

🔧 Customising Personas

Panelist personas are defined in the agents list inside main(). Edit the persona string to change how each model behaves:

Panelist(
    name="GPT-4o",
    provider="OpenAI",
    api_key=openai_key,
    persona="You are concise, logical, and academic. You focus on data and structural solutions.",
    color_class="agent-openai"
)

You can also swap in different model versions by editing the model strings inside the speak() method of the Panelist class.


🛠️ Extending the App

Some ideas for taking this further:

  • Auto-run mode — add a button that cycles through all three agents automatically with a configurable delay
  • Export transcript — add a download button to save the full discussion as Markdown or PDF
  • More panelists — add a fourth agent (e.g. Mistral, LLaMA via Ollama) by creating another Panelist instance
  • Smarter summarisation — replace the sentence-truncation summarize_text() with a real LLM call (GPT-3.5-Turbo or Gemini Flash are cheap and fast for this)
  • Voting / scoring — let the human moderator rate each response

📄 License

MIT — do whatever you like with it.

Ver o original no GitHub

Bom para

Relacionados