metorial-togetherai 1.0.0rc1

Together AI provider for Metorial

metorial-togetherai

Together AI provider integration for Metorial - enables using Metorial tools with Together AI's language models through OpenAI-compatible function calling.

Installation

pip install metorial-togetherai
# or
uv add metorial-togetherai
# or
poetry add metorial-togetherai

Features

• 🤖 Together AI Integration: Full support for Llama, Mixtral, and other Together AI models
• 🛠️ Function Calling: OpenAI-compatible function calling support
• 📡 Session Management: Automatic tool lifecycle handling
• 🔄 Format Conversion: Converts Metorial tools to OpenAI function format
• ⚡ Async Support: Full async/await support

Usage

Basic Usage

import asyncio
from openai import OpenAI
from metorial import Metorial
from metorial_togetherai import MetorialTogetherAISession

async def main():
    # Initialize clients
    metorial = Metorial(api_key="your-metorial-api-key")
    
    # Together AI uses OpenAI-compatible client
    together_client = OpenAI(
        api_key="your-together-api-key",
        base_url="https://api.together.xyz/v1"
    )
    
    # Create session with your server deployments
    async with metorial.session(["your-server-deployment-id"]) as session:
        # Create Together AI-specific wrapper
        together_session = MetorialTogetherAISession(session.tool_manager)
        
        messages = [
            {"role": "user", "content": "Help me with this task"}
        ]
        
        response = together_client.chat.completions.create(
            model="meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo",
            messages=messages,
            tools=together_session.tools
        )
        
        # Handle tool calls
        tool_calls = response.choices[0].message.tool_calls
        if tool_calls:
            tool_responses = await together_session.call_tools(tool_calls)
            
            # Add to conversation
            messages.append({
                "role": "assistant",
                "tool_calls": tool_calls
            })
            messages.extend(tool_responses)
            
            # Continue conversation...

asyncio.run(main())

Using Convenience Functions

from metorial_togetherai import build_togetherai_tools, call_togetherai_tools

async def example_with_functions():
    # Get tools in Together AI format
    tools = build_togetherai_tools(tool_manager)
    
    # Call tools from Together AI response
    tool_messages = await call_togetherai_tools(tool_manager, tool_calls)

API Reference

MetorialTogetherAISession

Main session class for Together AI integration.

session = MetorialTogetherAISession(tool_manager)

Properties:

• tools: List of tools in OpenAI-compatible format

Methods:

• async call_tools(tool_calls): Execute tool calls and return tool messages

build_togetherai_tools(tool_mgr)

Build Together AI-compatible tool definitions.

Returns: List of tool definitions in OpenAI format

call_togetherai_tools(tool_mgr, tool_calls)

Execute tool calls from Together AI response.

Returns: List of tool messages

Tool Format

Tools are converted to OpenAI-compatible format (without strict mode):

{
    "type": "function",
    "function": {
        "name": "tool_name",
        "description": "Tool description",
        "parameters": {
            "type": "object",
            "properties": {...},
            "required": [...]
        }
    }
}

Together AI API Configuration

Together AI uses the OpenAI-compatible API format. Configure your client like this:

from openai import OpenAI

client = OpenAI(
    api_key="your-together-api-key",
    base_url="https://api.together.xyz/v1"
)

Supported Models

Popular models available through Together AI:

• meta-llama/Meta-Llama-3.1-70B-Instruct-Turbo: Llama 3.1 70B
• meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo: Llama 3.1 8B
• mistralai/Mixtral-8x7B-Instruct-v0.1: Mixtral 8x7B
• NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO: Nous Hermes 2
• And many more...

Error Handling

try:
    tool_messages = await together_session.call_tools(tool_calls)
except Exception as e:
    print(f"Tool execution failed: {e}")

Tool errors are returned as tool messages with error content.

Dependencies

• metorial-openai-compatible>=1.0.0
• metorial-mcp-session>=1.0.0
• typing-extensions>=4.0.0

License

MIT License - see LICENSE file for details.