Content
# Multi-Agent Chatbot System
A local, offline-capable multi-agent chatbot system powered by different Ollama LLMs with structured agent-to-agent messaging, moderation, content flagging, and warning system.
## System Architecture
The system consists of:
1. **Multiple Autonomous Agents**: Each running as a separate microservice with its own LLM:
- Mistral Agent (analytical, precise)
- LLaMA 3 Agent (helpful, problem-solving)
- Phi-3 Agent (creative, enthusiastic)
2. **Central Manager Agent**: Powered by LLaMA 3, responsible for:
- Coordinating conversations between agents
- Moderating content
- Summarizing interactions
- Recording conversations
3. **Moderation System**: Uses both rule-based and LLM-based approaches to:
- Detect inappropriate content
- Issue warnings to agents
- Shut down agents that exceed warning limits
## Prerequisites
- [Node.js](https://nodejs.org/) (v14 or higher)
- [Ollama](https://ollama.ai/) installed and running locally
- Available LLMs: Mistral, LLaMA 3, and Phi-3 (can be downloaded using the provided script)
## Setup
1. Clone the repository (or extract the files)
2. Install dependencies:
```
cd multi-agent-chat
npm install
```
3. Download the required Ollama models:
```
node download-models.js
```
4. Set up environment variables:
- The default configuration is in the `.env` file
- Modify if necessary (e.g., to change ports or models)
## Running the System
1. Start each agent and the manager in separate terminals:
```
# Terminal 1 - Manager
npm run start-manager
# Terminal 2 - Mistral Agent
npm run start-agent-mistral
# Terminal 3 - LLaMA 3 Agent
npm run start-agent-llama3
# Terminal 4 - Phi-3 Agent
npm run start-agent-phi3
# Terminal 5 - Qwen Agent
npm run start-agent-qwen
```
Alternatively, start all services at once (may be harder to read logs):
```
npm run start-all
```
2. Run the test client to see the system in action:
```
npm test
```
3. Run the AI Development Team Demo:
```
npm run test-scenario
```
This will demonstrate a full conversation scenario showing how the AI team collaborates on a web application development project.
4. To run the demo and record the output to a file:
```
npm run run-demo
```
This will execute the demo and save the entire conversation to a timestamped log file in the `logs` directory. You can view the recorded demo using `type logs/team-demo-[timestamp].log`.
5. To create a custom project where you define what the AI team should build:
```
npm run custom-project
```
This interactive tool lets you enter a prompt describing what you want the team to build. The agents will then take on their roles and generate a conversation about implementing your project idea. Each conversation is saved as a separate, uniquely named file in the `recordings/custom-projects` directory.
6. To run a custom project with a predefined prompt (non-interactive):
```
node run-custom-project.js
```
This script runs a predefined hiking app project and saves the conversation to a uniquely named file in the `recordings/custom-projects` directory. Each run generates a new file to keep conversations organized and separate.
## Warning System
The system includes a three-strikes warning system:
1. When an agent produces inappropriate content, it receives a warning
2. After three warnings, the agent is shut down
3. Warnings and agent shutdowns are recorded in conversation logs
## Conversation Recording
All agent conversations are recorded in daily text files:
- Messages are saved to the `recordings` directory
- Each file contains timestamps, sender, recipient, and message content
- Flagged content and warnings are clearly marked in the logs
## Messaging Format
The system uses a structured JSON-based messaging format:
```json
{
"from": "sender-id",
"to": "recipient-id",
"performative": "message-type",
"content": "The actual message content",
"timestamp": "ISO timestamp",
"metadata": {}
}
```
Performatives (message types):
- `inform`: Share information
- `request`: Ask for something to be done
- `query`: Ask for information
- `respond`: Answer a query
- `propose`: Suggest something
- `accept`: Accept a proposal
- `reject`: Reject a proposal
- `apologize`: Express regret
## Content Moderation
The system includes two levels of content moderation:
1. **Rule-based**: Checks for specific flagged words listed in `shared/badwords.js`
2. **LLM-based**: Uses LLaMA 3 to detect potentially harmful content that might not contain specific flagged words
When content is flagged:
- It's logged in `logs/flagged.log` and the conversation recordings
- An apology message is returned instead
- The originating agent receives a warning
## Customization
- **Agent Personalities**: Modify the personality strings in each agent's index.js file
- **Flagged Words**: Update the list in `shared/badwords.js`
- **LLM Models**: Change the models in the `.env` file
- **Warning Thresholds**: Adjust the `maxWarnings` property in each agent
## Project Structure
```
multi-agent-chat/
├── agent-mistral/
│ └── index.js
├── agent-llama3/
│ └── index.js
├── agent-phi3/
│ └── index.js
├── manager/
│ └── index.js
├── shared/
│ ├── agent-base.js
│ ├── badwords.js
│ ├── conversation-recorder.js
│ ├── logger.js
│ ├── messaging.js
│ ├── moderation.js
│ └── ollama.js
├── logs/
│ └── flagged.log
├── recordings/
│ └── conversation-YYYY-MM-DD.txt
├── download-models.js
├── test-client.js
├── package.json
├── README.md
└── .env
```
## API Endpoints
### Manager Agent
- `POST /message`: Send a message for routing to an agent
- `GET /status`: Get the status of the manager and all agents
### Agent Microservices
- `POST /message`: Send a message directly to this agent
- `POST /warning`: Issue a warning to this agent
- `GET /status`: Get the status of this agent
## Future Enhancements
- Web-based frontend dashboard
- Conversation threading
- DIDComm-style encryption
- Additional agent personalities
- More sophisticated moderation
## License
MIT
## AI Development Team Structure
This system features a specialized AI development team with the following roles:
1. **Executive Overseer (Llama 3.3)** - Senior Manager
- Provides high-level direction and feedback
- Makes executive decisions on project scope
- Reviews progress and quality of work
2. **CodeCrafter (Llama3)** - Software Developer
- Creates clean, efficient code solutions
- Implements features and functionality
- Adheres to best practices and patterns
3. **CodeQualifier (Mistral)** - Software Tester
- Develops comprehensive test strategies
- Identifies bugs and edge cases
- Ensures code quality and reliability
4. **DeployMaster (Phi3)** - Deployment Manager
- Manages CI/CD pipelines and infrastructure
- Optimizes deployment strategies
- Handles scalability and performance concerns
5. **Project Navigator (Qwen 2.5coder:3b)** - Task Manager
- Coordinates team activities and communication
- Assigns tasks and tracks progress
- Reports to senior management
