debateEverything

Use agent framework to debate everything.

Project idea

This project explores a simple multi-agent debate workflow. A user gives a topic or task, then several agents respond in sequence:

1. Agent 1 gives an initial analysis or draft answer.
2. Agent 2 critiques Agent 1's response.
3. Agent 1 responds to the critique and improves the answer.
4. Agent 2 gives further feedback.
5. Agent 3 summarizes the whole discussion.

The goal is to use debate-style prompting to improve reasoning, critique, and final answers.

To-do notes

• debate_250416.py now uses ollama_api.py for the local Ollama /api/chat request format.
• ollama_api.py keeps Ollama as the default API and can also call an OpenAI-compatible /chat/completions API.
• Keep debate_dspy_250416.py as the currently smooth-running DSPy reference version.
• Previous code snapshots are stored in previous/ before the rewrite.
• Next: optimize the UI, either by creating a new version file or by using a more popular chat-style interface.

Archive

Previous versions of the main Python files are stored in previous/. These files are preserved before rewriting so old code is not deleted.

Version control

• debate_dspy_250416.py is for debating claims with DSPy.
• debate_code_250417.py is for debating code generation tasks.
• debate_250416.py is the rewritten Ollama/OpenAI-compatible API version.
• ollama_api.py contains the API helper used by debate_250416.py.

Files

• debate_250416.py — Streamlit UI for the rewritten debate workflow.
• ollama_api.py — API helper. Uses Ollama by default and can use an OpenAI-compatible API.
• debate_dspy_250416.py — DSPy + Ollama version. This has been the smooth-running reference version.
• debate_code_250417.py — code-generation debate workflow.
• previous/ — archived snapshots of earlier Python files.
• src/ — image assets used by the README.

Python environment

The project was checked with the Python environment recorded in the Claude/session environment:

D:\Users\niwakoki\miniconda3\envs\myenv\python.exe

Relevant packages found in that environment:

• streamlit
• requests

The rewritten Ollama version uses requests, so it does not require LangChain. The DSPy reference version requires dspy to be installed in the Python environment before running it.

How to run

1. Run the rewritten Ollama API debate version

This is the main rewritten version:

streamlit run debate_250416.py

Or run it with the recorded Python environment:

"D:\Users\niwakoki\miniconda3\envs\myenv\python.exe" -m streamlit run debate_250416.py

By default, this version opens with a clean sidebar. Detailed API settings are folded by default, and the default backend is local Ollama at:

http://localhost:11434/api/chat

Default model:

deepseek-r1:7b

You can change the API link, model name, Ollama format, API key/token, and temperature in the sidebar for the current session only. The sidebar also includes a Test API panel so you can send a short connection test before running a debate.

You can also override the startup defaults with environment variables:

OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=deepseek-r1:7b

Before running, make sure Ollama is open and the model exists locally. For example:

ollama list
ollama pull deepseek-r1:7b

2. Use an OpenAI-compatible API instead

ollama_api.py can also call an OpenAI-compatible /chat/completions API. Use this mode when connecting to a compatible local server or third-party endpoint.

The easiest way is to choose openai-compatible in the API settings panel after the app opens. Fill in the API link, model name, and optional API key/token there.

You can also set these environment variables before running the app if you want startup defaults:

DEBATE_API_MODE=openai
OPENAI_BASE_URL=http://localhost:1234/v1
OPENAI_MODEL=deepseek-r1:7b
OPENAI_API_KEY=your-key-if-needed

Then run:

streamlit run debate_250416.py

Notes:

• OPENAI_API_KEY can be empty if the local compatible server does not require a key.
• OPENAI_BASE_URL should point to the API root that contains /chat/completions.

3. Run the stable DSPy debate version

streamlit run debate_dspy_250416.py

This version uses DSPy and Ollama. Install DSPy first if your environment does not have it.

4. Run the code debate version

streamlit run debate_code_250417.py

This version asks agents to generate, critique, revise, and summarize code.

Saving debate results

After a debate finishes, the sidebar export panel provides download buttons for:

• Markdown (debate_session.md)
• HTML (debate_session.html)

The export includes the prompt, visible API mode/model information, temperature, and agent outputs. It does not include the API key/token.

Current API behavior

debate_250416.py shows API controls in the sidebar. Detailed API settings are collapsed by default. The user can choose:

• API mode: Ollama or OpenAI-compatible
• API link / base URL
• model name
• Ollama format: chat for /api/chat or generate for /api/generate
• API key / token for OpenAI-compatible APIs
• temperature

These UI settings are current-session only. They are stored in Streamlit session state, not written to files, and not committed to git. The API key/token field is a password field. A separate sidebar panel can test the selected API before starting a debate.

debate_250416.py calls:

get_ollama_response(prompt, role, temperature=0.7, api_settings=api_settings)

from ollama_api.py.

ollama_api.py then chooses the backend:

• default: Ollama native /api/chat
• optional: Ollama native /api/generate
• optional: OpenAI-compatible /chat/completions

Environment variables remain useful as startup defaults, but users can override them from the UI for the current session.

The active app keeps the debate workflow separate from API details.

UI optimization plan

The current UI is branded as Debate Everything and uses a cleaner Streamlit layout:

• API settings are in the sidebar and folded/collapsed by default.
• A separate Test API connection panel can check the selected backend before debate.
• The main page uses chat-style transcript messages for the user prompt and each agent response.
• Results are stored in st.session_state during the browser session.
• Completed debate results can be downloaded as Markdown or HTML.

Recommended next UI improvements:

1. Improve debate controls

   • Add a setting for the number of debate rounds.
   • Add a setting for custom agent names or roles.
   • Add optional streaming/progressive output if the backend supports it.

2. Possible popular UI directions

   • Short term: continue improving the current Streamlit app.
   • Later: Gradio ChatInterface for a simple public demo.
   • Later: React/Next.js chat frontend if the project needs a more polished web app.

Contribution

Credit to o3 Mini

AI-coding assist: gpt-4o, o3 Mini

References

Multiagent debate by autoGen

MAD

Tech stacks: streamlit, DSpy