spade-llm 0.2.0

Extension for SPADE to integrate Large Language Models in agents

Documentation | Quick Start | Examples | API Reference

SPADE-LLM: Large Language Model Integration for Multi-Agent Systems

SPADE-LLM is a Python framework that extends SPADE multi-agent platform with Large Language Model capabilities. Build AI agents powered by OpenAI GPT, Ollama, LM Studio, and other LLM providers for multi-agent systems, with communication via XMPP for distributed AI applications.

Keywords: SPADE, LLM, large language models, multi-agent systems, AI agents, OpenAI, GPT, Ollama, chatbot framework, distributed AI, Python AI, agent communication, XMPP agents, AI collaboration

Table of Contents

• Key Features
• Built-in XMPP Server
• Quick Start
• Installation
• Architecture
• Documentation
• Examples
  • Multi-Provider Support
  • Tools and Function Calling
  • Content Safety with Guardrails
  • Message Routing
  • Interactive Chat
  • Memory Extensions
  • Context Management
  • Human-in-the-Loop
• Requirements
• Contributing
• License

Key Features

• Built-in XMPP Server - No external server setup needed! Start agents with one command
• Multi-LLM Provider Support - Integrate OpenAI models, Ollama local models, LM Studio and more.
• Advanced Tool System - Function calling, async execution, LangChain tool integration, custom tool development
• Smart Context Management - Multi-conversation support, automatic cleanup, sliding window, token-aware context
• Persistent Memory - Agent-based memory, conversation threading, long-term state persistence across sessions
• Intelligent Message Routing - Conditional routing based on LLM responses, dynamic agent selection
• Content Safety Guardrails - Input/output filtering, keyword blocking, content moderation, safety controls
• MCP Integration - Model Context Protocol server support for external tools and services
• Human-in-the-Loop - Web interface for human expert consultation, interactive decision making

Built-in XMPP Server

SPADE 4+ includes a built-in XMPP server, eliminating the need for external server setup. This is a major advantage over other multi-agent frameworks like AutoGen or Swarm that require complex infrastructure configuration.

Start the Server

# Start SPADE's built-in XMPP server
spade run

The server automatically handles:

• Agent registration and authentication
• Message routing between agents
• Connection management
• Domain resolution

Agents automatically connect to the built-in server when using standard SPADE agent configuration.

Quick Start

Get started with SPADE-LLM in just 2 steps:

Step 1: Start the Built-in XMPP Server

# Terminal 1: Start SPADE's built-in server
spade run

Step 2: Create and Run Your LLM Agent

# your_agent.py
import spade
from spade_llm import LLMAgent, LLMProvider

async def main():
    provider = LLMProvider.create_openai(
        api_key="your-api-key",
        model="gpt-4o-mini"
    )
    
    agent = LLMAgent(
        jid="assistant@localhost",  # Connects to built-in server
        password="password",
        provider=provider,
        system_prompt="You are a helpful assistant"
    )
    
    await agent.start()

if __name__ == "__main__":
    spade.run(main())

# Terminal 2: Run your agent
python your_agent.py

Installation

pip install spade_llm

Examples

Multi-Provider Support

# OpenAI
provider = LLMProvider.create_openai(api_key="key", model="gpt-4o-mini")

# Ollama (local)
provider = LLMProvider.create_ollama(model="llama3.1:8b")

# LM Studio (local)
provider = LLMProvider.create_lm_studio(model="local-model")

Tools and Function Calling

from spade_llm import LLMTool

async def get_weather(city: str) -> str:
    return f"Weather in {city}: 22°C, sunny"

weather_tool = LLMTool(
    name="get_weather",
    description="Get weather for a city",
    parameters={
        "type": "object",
        "properties": {"city": {"type": "string"}},
        "required": ["city"]
    },
    func=get_weather
)

agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    tools=[weather_tool]
)

Content Safety with Guardrails

from spade_llm.guardrails import KeywordGuardrail, GuardrailAction

# Block harmful content
safety_filter = KeywordGuardrail(
    name="safety_filter",
    blocked_keywords=["hack", "exploit", "malware"],
    action=GuardrailAction.BLOCK,
    blocked_message="I cannot help with potentially harmful activities."
)

agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    input_guardrails=[safety_filter]  # Filter incoming messages
)

Message Routing

def router(msg, response, context):
    if "technical" in response.lower():
        return "[e-mail]"
    return str(msg.sender)

agent = LLMAgent(
    jid="router@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    routing_function=router
)

Interactive Chat

from spade_llm import ChatAgent

# Create chat interface
chat_agent = ChatAgent(
    jid="human@localhost",  # Uses built-in server
    password="password",
    target_agent_jid="assistant@localhost"
)

await chat_agent.start()
await chat_agent.run_interactive()  # Start interactive chat

Memory Extensions

# Agent-based memory: Single shared memory per agent
agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    agent_base_memory=(True, "./memory.db")  # Enabled with custom path
)

# Agent-thread memory: Separate memory per conversation
agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    agent_thread_memory=(True, "./thread_memory.db")  # Enabled with custom path
)

# Default memory paths (if path not specified)
agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    agent_base_memory=(True, None)  # Uses default path
)

Context Management

from spade_llm.context import SmartWindowSizeContext, FixedWindowSizeContext

# Smart context: Dynamic window sizing based on content
smart_context = SmartWindowSizeContext(
    max_tokens=4000,
    include_system_prompt=True,
    preserve_last_k_messages=5
)

# Fixed context: Traditional sliding window
fixed_context = FixedWindowSizeContext(
    max_messages=20,
    include_system_prompt=True
)

agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    context_manager=smart_context
)

Human-in-the-Loop

from spade_llm import HumanInTheLoopTool

# Create tool for human consultation
human_tool = HumanInTheLoopTool(
    human_expert_jid="expert@localhost",  # Uses built-in server
    timeout=300.0  # 5 minutes
)

agent = LLMAgent(
    jid="assistant@localhost",  # Uses built-in server
    password="password",
    provider=provider,
    tools=[human_tool]  # Pass tools in constructor
)

# Start web interface for human expert
# python -m spade_llm.human_interface.web_server
# Open http://localhost:8080 and connect as expert

Architecture

graph LR
    A[LLMAgent] --> C[ContextManager]
    A --> D[LLMProvider]
    A --> E[LLMTool]
    A --> G[Guardrails]
    A --> M[Memory]
    D --> F[OpenAI/Ollama/etc]
    G --> H[Input/Output Filtering]
    E --> I[Human-in-the-Loop]
    E --> J[MCP]
    E --> P[CustomTool/LangchainTool]
    J --> K[STDIO]
    J --> L[HTTP Streaming]
    M --> N[Agent-based]
    M --> O[Agent-thread]

Documentation

• Installation - Setup and requirements
• Quick Start - Basic usage examples
• Providers - LLM provider configuration
• Tools - Function calling system
• Guardrails - Content filtering and safety
• API Reference - Complete API documentation

Examples Directory

The /examples directory contains complete working examples:

• multi_provider_chat_example.py - Chat with different LLM providers
• ollama_with_tools_example.py - Local models with tool calling
• langchain_tools_example.py - LangChain tool integration
• guardrails_example.py - Content filtering and safety controls
• human_in_the_loop_example.py - Human expert consultation via web interface
• valencia_multiagent_trip_planner.py - Multi-agent workflow

Requirements

• Python 3.10+
• SPADE 3.3.0+

Contributing

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Submit a pull request

See Contributing Guide for details.

License

MIT License