self-constructive-MAS-for-process-development

MAS is a Streamlit-based web platform for testing and visualizing Multi-Agent Systems (MAS) designed for autonomous chemical reasoning and workflow execution.

---

🚀 Quick Start (Docker Deployment)

1️⃣ Build the image

docker build -t mas:v0.1 .

2️⃣ Run the container

docker run -d   --name streamlit   -e TZ=Asia/Shanghai LLM_BASE_URL=xxx -e LLM_BASE_API_KEY=xxx -p 8501:8501   mas:v0.1

🕒 TZ=Asia/Shanghai ensures the app runs in Beijing Time (UTC +8).
After startup, open http://localhost:8501 in your browser.

---

📁 Project Structure

mas/
├── assets/                   # Static assets (images, icons, etc.)
├── config/                   # MAS configuration files (MAS / tools JSON)
├── data/                     # General data directory
├── flow/                     # Flow logs
├── history/                  # Execution history logs
├── memory/                   # Agent memory storage
├── pages/                    # Streamlit page modules
│   ├── 1-LLM_config.py
│   ├── 2-Tools_box.py
│   ├── 3-Multi-agent_generation.py
│   ├── 4-Playground.py
│   └── 5-History.py
│
├── agent.py                  # Agent class definitions
├── Dockerfile                # Docker build configuration
├── flow.py                   # Main flow control
├── functions.py              # Tool function implementations
├── Home.py                   # Streamlit entry page
├── llm.py                    # LLM class definitions
├── manager.py                # Manager agent
├── requirements.txt          # Python dependencies
├── run.py                    # Workflow runner
├── task.py                   # Task class definitions
├── template.txt              # Prompt or template base file
└── worker.py                 # Worker agent logic

---

---

🖥️ Application Usage Guide

Once the container is running, open:
👉 http://localhost:8501

The MAS-Playground interface contains five major pages, each serving a specific role in the multi-agent workflow:

---

🧠 1. LLM Config

This page allows you to configure and test the language model backend used by the agents.

• Define the model name.
• Set API key, base URL, and version.
• All settings are saved into config/llm_config.json for later use by the MAS.

---

🧰 2. Tools Box

A toolbox for managing and editing all available agent tools.

• Each tool consists of:
  • Function name and natural-language description.
  • Python code implementation.
  • Automatically generated JSON Schema for worker agent calling.
• The page can:
  • Auto-generate JSON schema from the Python code body.
  • Allow users to edit or delete tools dynamically.
  • Display scrollable code blocks and structured metadata for each tool.
• All tools are stored in config/tools_config.json and are dynamically bound to worker agents in runtime.

---

🤖 3. Multi-Agent Generation

A page for self-generating, visualizing, and self-modifying MAS architectures.

🔹 Automatic Generation

• Supports automatic MAS generation based on predefined templates or natural-language prompts.
• Automatically creates manager–worker–task relationships and binds each worker to relevant tools.
• Saves the generated MAS configuration as config/mas_config.json for later use in the Playground.

🔹 Self-Modification

• Enables self-modification of existing MAS configurations.
• Changes can be applied to selected sections without rebuilding the entire MAS.

🔹 Visualization and Inspection

• Displays the list of existing MAS architectures saved in the project.
• For a selected MAS, clicking Manager, Workers, Tasks,or Tools reveals the exact JSON configuration of that node (parameters, bindings, and connections).

---

🧩 4. Playground

This is the interactive execution and visualization core of the system.

• Select one MAS configuration to launch.
• In the Conversation tab:
  • Type a prompt or instruction for the MAS (e.g., “Find the possible synthesis procedure for X”).
  • Optionally upload a related file (PDF, TXT, or MD), which will be saved under /app/data/uploads/.
  • The system will create a unique flow_id and start the reasoning workflow.
  • Each reasoning step, including worker/tool interactions, is displayed in real time.
• In the Flow Chart tab:
  • View the MAS topology (manager → task → worker → tool connections).
  • Monitor live data flow and reasoning sequence updates.

---

📜 5. History

A record and review page for past workflow executions.

• All completed workflows and memory logs are stored under /memory/ and /flow/.
• You can reload previous flow_ids to inspect results and agent reasoning traces.

---

🧠 Example Workflow

1. Go to LLM Config to set up the desired language model.
2. Open Tools Box to ensure the correct functions are available to the agents.
3. Use Multi-Agent Generation to create a new MAS configuration or modify an existing one.
4. Launch Playground, select your MAS, and input a research query.
5. Observe real-time reasoning, intermediate outputs, and visual flow in the Flow Chart tab.
6. After execution, switch to History to revisit or analyze previous results.

---