rakam-systems-agent 0.1.1rc11

AI Agents framework with Pydantic AI support and LLM Gateway integration

Rakam System Agent

The agent package of Rakam Systems providing AI agent implementations powered by Pydantic AI.

Overview

rakam-systems-agent provides flexible AI agents with tool integration, chat history, and LLM gateway abstractions. This package depends on rakam-systems-core.

Features

• Configuration-First Design: Change agents without code changes - just update YAML files
• Async/Sync Support: Full support for both synchronous and asynchronous agent operations
• Tool Integration: Easy tool definition and integration using the Tool.from_schema pattern
• Model Settings: Control model behavior including parallel tool calls, temperature, and max tokens
• Pydantic AI Powered: Built on top of Pydantic AI library
• Streaming Support: Both sync and async streaming interfaces
• Chat History: Multiple backends (JSON, SQLite, PostgreSQL)
• LLM Gateway: Unified interface for OpenAI and Mistral AI

🎯 Configuration Convenience

The agent package supports comprehensive YAML configuration, allowing you to:

• Switch LLM models without changing code (GPT-4 → GPT-4o-mini → Claude)
• Tune parameters instantly (temperature, max tokens, parallel tools)
• Modify prompts without redeployment
• Add/remove tools by editing config files
• Enable/disable tracking via configuration
• Use different configs for different environments

Example: Switch from GPT-4o to GPT-4o-mini by changing one line in your YAML - no code changes, no redeployment needed!

Installation

# Requires core package
pip install -e ./rakam-systems-core

# Install agent package
pip install -e ./rakam-systems-agent

Quick Start

Using BaseAgent (Pydantic AI-powered)

import asyncio
from rakam_system_agent import BaseAgent
from rakam_systems_core.interfaces import ModelSettings
from rakam_systems_core.interfaces.tool import ToolComponent as Tool

# Define a tool function
async def get_weather(city: str) -> dict:
    """Get weather information for a city"""
    # Your implementation here
    return {"city": city, "temperature": 72, "condition": "sunny"}

# Create an agent with tools
agent = BaseAgent(
    name="weather_agent",
    model="openai:gpt-4o",
    system_prompt="You are a helpful weather assistant.",
    tools=[
        Tool.from_schema(
            function=get_weather,
            name='get_weather',
            description='Get weather information for a city',
            json_schema={
                'type': 'object',
                'properties': {
                    'city': {'type': 'string', 'description': 'The city name'},
                },
                'required': ['city'],
                'additionalProperties': False,
            },
            takes_ctx=False,
        ),
    ],
)

# Run the agent
async def main():
    result = await agent.arun(
        "What's the weather in San Francisco?",
        model_settings=ModelSettings(parallel_tool_calls=True),
    )
    print(result.output_text)

asyncio.run(main())

Core Components

AgentComponent

The base abstract class for all agents. Provides:

• run() / arun(): Execute the agent synchronously or asynchronously
• stream() / astream(): Stream responses
• Support for tools, model settings, and dependencies

BaseAgent

The core agent implementation powered by Pydantic AI. This is the primary agent class in our system. Features:

• Direct integration with Pydantic AI's Agent
• Full support for parallel tool calls
• Automatic conversion between our interfaces and Pydantic AI's
• Support for both traditional tool lists and ToolRegistry/ToolInvoker system

Tool

Wrapper for tool functions compatible with Pydantic AI's Tool.from_schema pattern:

Tool.from_schema(
    function=my_function,
    name='my_function',
    description='What this function does',
    json_schema={
        'type': 'object',
        'properties': {...},
        'required': [...],
    },
    takes_ctx=False,
)

ModelSettings

Configure model behavior:

ModelSettings(
    parallel_tool_calls=True,  # Enable parallel tool execution
    temperature=0.7,           # Control randomness
    max_tokens=1000,           # Limit response length
)

Advanced Usage

Parallel vs Sequential Tool Calls

Control whether tools are called in parallel or sequentially:

# Parallel (faster for independent tools)
result = await agent.arun(
    "Get weather for NYC and LA",
    model_settings=ModelSettings(parallel_tool_calls=True),
)

# Sequential (for dependent operations)
result = await agent.arun(
    "Get weather for NYC and LA",
    model_settings=ModelSettings(parallel_tool_calls=False),
)

Using Dependencies

Pass context/dependencies to your agent:

class Deps:
    def __init__(self, user_id: str):
        self.user_id = user_id

agent = BaseAgent(
    deps_type=Deps,
    # ...
)

result = await agent.arun(
    "Process this",
    deps=Deps(user_id="123"),
)

Streaming Responses

async for chunk in agent.astream("Tell me a story"):
    print(chunk, end='', flush=True)

API Reference

AgentComponent

class AgentComponent(BaseComponent):
    def __init__(
        self,
        name: str,
        config: Optional[Dict[str, Any]] = None,
        model: Optional[str] = None,
        deps_type: Optional[Type[Any]] = None,
        system_prompt: Optional[str] = None,
        tools: Optional[List[Any]] = None,
    )

    def run(
        self,
        input_data: Union[str, AgentInput],
        deps: Optional[Any] = None,
        model_settings: Optional[ModelSettings] = None
    ) -> AgentOutput

    async def arun(
        self,
        input_data: Union[str, AgentInput],
        deps: Optional[Any] = None,
        model_settings: Optional[ModelSettings] = None
    ) -> AgentOutput

Tool

class Tool:
    @classmethod
    def from_schema(
        cls,
        function: Callable[..., Any],
        name: str,
        description: str,
        json_schema: Dict[str, Any],
        takes_ctx: bool = False,
    ) -> "Tool"

ModelSettings

class ModelSettings:
    def __init__(
        self,
        parallel_tool_calls: bool = True,
        temperature: Optional[float] = None,
        max_tokens: Optional[int] = None,
        **kwargs: Any
    )

Examples

See the examples/ai_agents_examples/ directory in the main repository for complete examples demonstrating:

• Multiple tool definitions
• Parallel vs sequential tool calls
• Performance comparisons
• Complex multi-tool workflows
• Chat history integration
• RAG systems

Package Structure

rakam-systems-agent/
├── src/rakam_system_agent/
│   ├── components/
│   │   ├── base_agent.py         # BaseAgent (Pydantic AI-powered)
│   │   ├── llm_gateway/          # LLM provider gateways
│   │   ├── chat_history/         # Chat history backends
│   │   ├── tools/                # Built-in tools
│   │   └── __init__.py           # Exports
│   └── server/                   # MCP server
└── pyproject.toml

Best Practices

1. Use async when possible: Async operations are more efficient, especially with tools
2. Enable parallel tool calls: For independent operations, parallel execution is much faster
3. Provide clear tool descriptions: Better descriptions help the LLM use tools correctly
4. Use type hints: JSON schemas should match your function signatures
5. Handle errors gracefully: Tools should catch and return meaningful errors

Migration Guide

Using BaseAgent

BaseAgent is the core Pydantic AI-powered agent implementation in this framework.

from rakam_system_agent import BaseAgent

agent = BaseAgent(
    name="agent",
    model="openai:gpt-4o",
    system_prompt="You are a helpful assistant."
)

Troubleshooting

ImportError: pydantic_ai not installed

Install Pydantic AI:

pip install pydantic_ai

Tool not being called

Check:

1. Tool description is clear and relevant
2. JSON schema matches function signature
3. System prompt doesn't contradict tool usage

Performance issues

• Enable parallel_tool_calls=True for independent operations
• Use async functions for I/O-bound operations
• Consider caching tool results when appropriate

Contributing

When adding new agent types:

1. Inherit from BaseAgent
2. Implement ainfer() for async or infer() for sync
3. Add tests in tests/
4. Update this README with examples

License

See main project LICENSE file.