multi-llm-orchestrator 0.6.0

Unified interface for Russian LLMs with intelligent routing and fallback

Multi-LLM Orchestrator

Architecture

Multi-LLM Orchestrator provides automatic failover between GigaChat, YandexGPT, and Ollama with streaming support.

A unified interface for orchestrating multiple Large Language Model providers with intelligent routing and fallback mechanisms.

Overview

The Multi-LLM Orchestrator provides a seamless way to integrate and manage multiple LLM providers through a single, consistent interface. It supports intelligent routing strategies, automatic fallbacks, provider-level metrics tracking, and provider-specific optimizations. Currently focused on Russian LLM providers (GigaChat, YandexGPT) with a flexible architecture that supports any LLM provider implementation.

Features

• Multiple LLM Providers: Unified interface for GigaChat, YandexGPT, Ollama, and custom providers
• Intelligent Routing: Multiple routing strategies including round-robin, random, first-available, and best-available (health + latency aware)
• Automatic Fallback: Seamless failover when providers fail
• Provider-level Metrics: Track latency, success/failure rates, and health status for each provider
• Smart Routing Strategy: best-available strategy selects the healthiest provider with lowest latency based on real-time metrics
• Streaming Support: Incremental text generation with streaming responses
• LangChain Integration: Optional compatibility layer for LangChain chains and prompts

Quickstart

Get started with Multi-LLM Orchestrator in minutes:

Using MockProvider (Testing)

import asyncio
from orchestrator import Router
from orchestrator.providers import ProviderConfig, MockProvider

async def main():
    # Initialize router with round-robin strategy
    router = Router(strategy="round-robin")
    
    # Add providers
    for i in range(3):
        config = ProviderConfig(name=f"provider-{i+1}", model="mock-normal")
        router.add_provider(MockProvider(config))
    
    # Make a request
    response = await router.route("What is Python?")
    print(response)
    # Output: Mock response to: What is Python?

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

Using GigaChatProvider (Production)

import asyncio
from orchestrator import Router
from orchestrator.providers import ProviderConfig, GigaChatProvider

async def main():
    # Create GigaChat provider
    config = ProviderConfig(
        name="gigachat",
        api_key="your_authorization_key_here",  # OAuth2 authorization key
        model="GigaChat",  # or "GigaChat-Pro", "GigaChat-Plus"
        scope="GIGACHAT_API_PERS"  # or "GIGACHAT_API_CORP" for corporate
    )
    provider = GigaChatProvider(config)
    
    # Use with router
    router = Router(strategy="round-robin")
    router.add_provider(provider)
    
    # Generate response
    response = await router.route("What is Python?")
    print(response)

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

Disabling SSL Verification (for self-signed certificates)

If you encounter SSL certificate errors with GigaChat (Russian CA certificates), you can disable verification:

import asyncio
from orchestrator import Router
from orchestrator.providers import GigaChatProvider, ProviderConfig

async def main():
    router = Router(strategy="round-robin")
    
    # WARNING: Disabling SSL verification is insecure
    # Use only in development or with trusted networks
    config = ProviderConfig(
        name="gigachat",
        api_key="your_authorization_key_here",
        scope="GIGACHAT_API_PERS",
        verify_ssl=False  # Disable SSL verification
    )
    
    router.add_provider(GigaChatProvider(config))
    
    response = await router.route("Hello!")
    print(response)

asyncio.run(main())

⚠️ Security Warning: Disabling SSL verification makes your application vulnerable to man-in-the-middle attacks. Use this option only in development or when working with known self-signed certificates.

Using YandexGPTProvider (Production)

import asyncio
from orchestrator import Router
from orchestrator.providers import ProviderConfig, YandexGPTProvider

async def main():
    # Create YandexGPT provider
    config = ProviderConfig(
        name="yandexgpt",
        api_key="your_iam_token_here",  # IAM token (valid for 12 hours)
        folder_id="your_folder_id_here",  # Yandex Cloud folder ID
        model="yandexgpt/latest"  # or "yandexgpt-lite/latest"
    )
    provider = YandexGPTProvider(config)
    
    # Use with router
    router = Router(strategy="round-robin")
    router.add_provider(provider)
    
    # Generate response
    response = await router.route("What is Python?")
    print(response)

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

Local Models with Ollama

Run open-source LLMs locally without API keys:

import asyncio
from orchestrator import Router
from orchestrator.providers import OllamaProvider, ProviderConfig

async def main():
    router = Router(strategy="first-available")

    ollama_config = ProviderConfig(
        name="ollama",
        model="llama3",  # or "mistral", "phi", etc.
        base_url="http://localhost:11434",  # optional; defaults to localhost
    )
    router.add_provider(OllamaProvider(ollama_config))

    response = await router.route("Why is the sky blue?")
    print(response)

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

Requirements: Install Ollama from https://ollama.ai and pull a model (e.g., ollama pull llama3).

The MockProvider simulates LLM behavior without requiring API credentials, while GigaChatProvider and YandexGPTProvider provide full integration with their respective APIs.

Installation

Requirements:

• Python 3.11+
• Poetry (recommended) or pip

Using Poetry

# Clone the repository
git clone https://github.com/MikhailMalorod/Multi-LLM-Orchestrator.git
cd Multi-LLM-Orchestrator

# Install dependencies
poetry install

Using pip

# Clone the repository
git clone https://github.com/MikhailMalorod/Multi-LLM-Orchestrator.git
cd Multi-LLM-Orchestrator

# Install in development mode
pip install -e .

Architecture

The Multi-LLM Orchestrator follows a modular architecture with clear separation of concerns:

┌──────────────────────────────────────────────┐
│              User Application                │
└─────────────────┬────────────────────────────┘
                  │
                  ▼
         ┌────────────────┐
         │     Router     │ ◄── Strategy: round-robin/random/first-available/best-available
         └────────┬───────┘
                  │
      ┌───────────┼───────────┐
      ▼           ▼           ▼
┌──────────┐ ┌──────────┐ ┌──────────┐
│Provider 1│ │Provider 2│ │Provider 3│
│(Base)    │ │(Base)    │ │(Base)    │
└────┬─────┘ └────┬─────┘ └────┬─────┘
     │            │            │
     ▼            ▼            ▼
   (API)        (API)        (API)

Components

• Router (src/orchestrator/router.py): Manages provider selection based on routing strategy and handles automatic fallback when providers fail.

• BaseProvider (src/orchestrator/providers/base.py): Abstract base class defining the interface that all provider implementations must follow. Includes configuration models (ProviderConfig, GenerationParams) and exception hierarchy.

• MockProvider (src/orchestrator/providers/mock.py): Test implementation that simulates LLM behavior without making actual API calls. Supports various simulation modes for testing different scenarios.

• Config (src/orchestrator/config.py): Future component for loading configuration from environment variables. Currently used for planned real provider integrations (GigaChat, YandexGPT).

Routing Strategies

The Router supports four routing strategies, each suitable for different use cases:

Strategy	Description	Use Case
round-robin	Cycles through providers in a fixed order	Equal load distribution (recommended for production)
random	Selects a random provider from available providers	Simple random selection for load balancing
first-available	Selects the first healthy provider based on health checks	High availability scenarios with automatic unhealthy provider skipping
best-available	Selects the healthiest provider with lowest latency based on real-time metrics	Production environments requiring optimal performance and reliability

The strategy is selected when initializing the Router:

router = Router(strategy="round-robin")  # or "random", "first-available", or "best-available"

Best-Available Strategy

The best-available strategy uses provider health status and latency metrics to intelligently route requests:

• Health Status: Providers are categorized as healthy, degraded, or unhealthy based on error rates and latency patterns
• Latency Optimization: Among providers with the same health status, selects the one with the lowest rolling average latency
• Automatic Adaptation: Metrics are updated in real-time, so routing decisions adapt as provider performance changes

import asyncio
from orchestrator import Router
from orchestrator.providers import ProviderConfig, GigaChatProvider, YandexGPTProvider

async def main():
    # Initialize router with best-available strategy
    router = Router(strategy="best-available")
    
    # Add multiple providers
    router.add_provider(GigaChatProvider(ProviderConfig(
        name="gigachat", api_key="key1", model="GigaChat"
    )))
    router.add_provider(YandexGPTProvider(ProviderConfig(
        name="yandexgpt", api_key="key2", folder_id="folder1", model="yandexgpt/latest"
    )))
    
    # Router will automatically select the healthiest and fastest provider
    response = await router.route("What is Python?")
    print(response)

asyncio.run(main())

The Router tracks performance metrics for each provider (latency, success rate, error rate) and uses this data to make intelligent routing decisions. Providers with high error rates or degraded latency are automatically deprioritized.

Run the Demo

See the routing strategies and fallback mechanisms in action:

python examples/routing_demo.py

No API keys required — uses MockProvider for demonstration.

The demo showcases:

• All four routing strategies (round-robin, random, first-available, best-available)
• Automatic fallback mechanism when providers fail
• Error handling when all providers are unavailable

See routing_demo.py for the complete interactive demonstration.

MockProvider Modes

MockProvider simulates various LLM behaviors for testing without requiring API credentials:

• mock-normal — Returns successful responses with a small delay
• mock-timeout — Simulates timeout errors
• mock-unhealthy — Health check returns False (useful for testing first-available strategy)
• mock-ratelimit — Simulates rate limit errors
• mock-auth-error — Simulates authentication failures

See mock.py for all available modes and detailed documentation.

Roadmap

See our GitHub Issues for planned features and roadmap updates.

Current Status

• ✅ Core architecture with Router and BaseProvider
• ✅ MockProvider for testing
• ✅ GigaChatProvider with OAuth2 authentication
• ✅ Four routing strategies (round-robin, random, first-available, best-available)
• ✅ Provider-level metrics tracking (latency, success/failure, health status)
• ✅ Automatic fallback mechanism
• ✅ Example demonstrations

Supported Providers

• ✅ MockProvider — For testing and development
• ✅ GigaChatProvider — Full integration with GigaChat (Sber) API
  • OAuth2 authentication with automatic token refresh
  • Support for all generation parameters
  • Comprehensive error handling
• ✅ YandexGPTProvider — Full integration with YandexGPT (Yandex Cloud) API
  • IAM token authentication (user-managed, 12-hour validity)
  • Support for temperature and maxTokens parameters
  • Support for yandexgpt/latest and yandexgpt-lite/latest models
  • Comprehensive error handling
• ✅ OllamaProvider — Local models (Llama 3, Mistral, Phi) via Ollama API

Planned Providers

• Additional open-source providers (TBD)

LangChain Integration

Note: Requires optional dependency. Install with:

pip install multi-llm-orchestrator[langchain]

Use Multi-LLM Orchestrator providers with LangChain chains, prompts, and other LangChain components:

from langchain_core.prompts import ChatPromptTemplate
from orchestrator.langchain import MultiLLMOrchestrator
from orchestrator import Router
from orchestrator.providers import GigaChatProvider, ProviderConfig

# Create router with providers
router = Router(strategy="round-robin")
config = ProviderConfig(
    name="gigachat",
    api_key="your_api_key",
    model="GigaChat"
)
router.add_provider(GigaChatProvider(config))

# Use as LangChain LLM
llm = MultiLLMOrchestrator(router=router)

# Work with LangChain chains
prompt = ChatPromptTemplate.from_template("Tell me about {topic}")
chain = prompt | llm
response = chain.invoke({"topic": "Python"})

The MultiLLMOrchestrator class implements LangChain's BaseLLM interface, supporting both synchronous and asynchronous calls. All routing strategies and fallback mechanisms work seamlessly with LangChain.

Streaming Support

Multi-LLM Orchestrator now supports streaming responses, allowing you to receive text chunks incrementally as they are generated. This is especially useful for real-time applications and improved user experience.

Basic Streaming with Router

import asyncio
from orchestrator import Router
from orchestrator.providers import ProviderConfig, MockProvider

async def main():
    router = Router(strategy="round-robin")
    config = ProviderConfig(name="mock", model="mock-normal")
    router.add_provider(MockProvider(config))
    
    # Stream response chunk by chunk
    async for chunk in router.route_stream("What is Python?"):
        print(chunk, end="", flush=True)

asyncio.run(main())

Streaming with LangChain

from orchestrator.langchain import MultiLLMOrchestrator
from orchestrator import Router
from orchestrator.providers import MockProvider, ProviderConfig

router = Router(strategy="round-robin")
router.add_provider(MockProvider(ProviderConfig(name="mock", model="mock-normal")))

llm = MultiLLMOrchestrator(router=router)

# Async streaming
async for chunk in llm._astream("What is Python?"):
    print(chunk, end="", flush=True)

# Sync streaming
for chunk in llm._stream("What is Python?"):
    print(chunk, end="", flush=True)

Streaming Features

• Incremental responses: Receive text chunks as they are generated
• Fallback support: Automatic provider fallback works before the first chunk is yielded
• Provider support: Currently supported in MockProvider and GigaChatProvider
• LangChain integration: Full support for both sync and async streaming in LangChain

Streaming Examples

See streaming_demo.py and langchain_streaming_demo.py for complete examples.

Provider Metrics & Monitoring

Multi-LLM Orchestrator automatically tracks performance metrics for each provider, enabling intelligent routing and monitoring.

Accessing Metrics

The Router collects aggregated metrics for each provider, including:

• Request counts (total, successful, failed)
• Average latency (for successful requests)
• Rolling average latency (last 100 requests)
• Error rate (recent errors)
• Health status (healthy, degraded, or unhealthy)

from orchestrator import Router
from orchestrator.providers import ProviderConfig, MockProvider

router = Router(strategy="best-available")
router.add_provider(MockProvider(ProviderConfig(name="provider1", model="mock-normal")))

# Make some requests
await router.route("Test 1")
await router.route("Test 2")

# Access metrics
metrics = router.get_metrics()
for provider_name, provider_metrics in metrics.items():
    print(f"{provider_name}:")
    print(f"  Health: {provider_metrics.health_status}")
    print(f"  Success rate: {provider_metrics.success_rate:.2%}")
    print(f"  Avg latency: {provider_metrics.avg_latency_ms:.1f}ms")
    print(f"  Rolling avg latency: {provider_metrics.rolling_avg_latency_ms:.1f}ms" if provider_metrics.rolling_avg_latency_ms else "  Rolling avg latency: N/A")

Health Status

Provider health status is determined automatically based on:

• Error Rate: High error rates (>30% degraded, >60% unhealthy) indicate provider issues
• Latency Degradation: If rolling average latency is significantly higher than overall average, provider is marked as degraded
• Insufficient Data: New providers with few requests are optimistically marked as healthy

The best-available routing strategy uses health status to prioritize providers, always preferring healthy over degraded over unhealthy.

Structured Logging

Router automatically logs request events with structured fields:

• llm_request_completed (info level) for successful requests
• llm_request_failed (warning level) for failed requests

Each log entry includes: provider, model, latency_ms, streaming, success, and error_type (for failures).

Note: Token-based metrics (token count, tokens/s, cost) are not yet implemented. This is planned for future releases (v0.7.0+).

Documentation

• Architecture Overview — System design and components
• Contributing Guide — How to contribute to the project
• Provider Documentation — Detailed provider guides
  • GigaChat Provider
  • YandexGPT Provider
  • Creating Custom Provider
• routing_demo.py — Interactive demonstration of routing strategies and fallback mechanisms

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/amazing-feature)
3. Commit your changes (git commit -m 'Add some amazing feature')
4. Push to the branch (git push origin feature/amazing-feature)
5. Open a Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.