langchain-fused-model 0.1.2

Intelligent routing and management for multiple LangChain ChatModel instances with rate limiting, fallback, and structured output support

langchain-fused-model

Intelligent routing and management for multiple LangChain ChatModel instances with advanced features like rate limiting, automatic fallback, and structured output support.

Table of Contents

• Overview
• Why langchain-fused-model
• Features
• Installation
• Quick Start
• Routing Strategies
  • Priority-Based Routing
  • Cost-Aware Routing
  • Round-Robin Routing
  • Least-Used Routing
  • Custom Strategies
• Structured Output
• Rate Limiting and Fallback
• LangChain Integration
• Usage Statistics
• Advanced Configuration
• Examples
• Requirements
• Contributing
• License
• Support

---

Overview

langchain-fused-model provides a MultiModelManager class that acts as a unified interface for managing multiple LangChain ChatModel instances. It enables dynamic model selection based on configurable strategies while maintaining full LangChain compatibility.

The manager inherits from LangChain's BaseChatModel, making it a drop-in replacement for any ChatModel in chains, agents, and other workflows.

Why langchain-fused-model

Many developers today rely on multiple large language model providers to balance cost, availability, latency, and capabilities. However, LangChain does not provide a unified interface to dynamically route across multiple models based on rate limits or priorities. This project was created to fill that gap.

Whether you're managing free-tier APIs, orchestrating across OpenAI and Anthropic, or experimenting with cost-based strategies, langchain-fused-model helps you:

• Fail gracefully when APIs are throttled or down
• Reduce latency or cost by routing requests optimally
• Extract structured outputs even from models that don't support it natively
• Scale production chains and agents with built-in observability and fallback

Features

• Multiple Routing Strategies: Priority-based, round-robin, least-used, and cost-aware routing
• Automatic Rate Limiting: Per-model rate limits (RPM/RPS) with automatic fallback
• Error Resilience: Automatic fallback to alternative models on failures
• Structured Output: Pydantic-validated responses with native support detection and fallback
• Full LangChain Compatibility: Implements BaseChatModel and Runnable interfaces
• Usage Tracking: Monitor requests, tokens, and success rates per model
• Extensible: Support for custom routing strategies and error handlers
• Production Ready: Comprehensive logging and error handling

Installation

Install from PyPI:

pip install langchain-fused-model

For development installation:

git clone https://github.com/yourusername/langchain-fused-model
cd langchain-fused-model
pip install -e .

Quick Start

Here's a simple example to get you started:

from langchain_openai import ChatOpenAI
from langchain_anthropic import ChatAnthropic
from langchain_fused_model import MultiModelManager, RoutingStrategy

# Initialize your models
models = [
    ChatOpenAI(model="gpt-4"),
    ChatOpenAI(model="gpt-3.5-turbo"),
    ChatAnthropic(model="claude-3-opus-20240229"),
]

# Create manager with priority-based routing
manager = MultiModelManager(
    models=models,
    strategy=RoutingStrategy.PRIORITY
)

# Use like any LangChain ChatModel
response = manager.invoke("What is the capital of France?")
print(response.content)

Routing Strategies

The MultiModelManager supports multiple routing strategies to control how requests are distributed across models.

Priority-Based Routing

Routes requests to the highest priority available model. Perfect for preferring premium models with fallback to cheaper alternatives.

from langchain_fused_model import MultiModelManager, RoutingStrategy, ModelConfig

configs = [
    ModelConfig(priority=100, max_rpm=60),  # Highest priority - GPT-4
    ModelConfig(priority=50, max_rpm=120),  # Medium priority - GPT-3.5
    ModelConfig(priority=10, max_rpm=200),  # Lowest priority - Local model
]

manager = MultiModelManager(
    models=models,
    model_configs=configs,
    strategy=RoutingStrategy.PRIORITY
)

Cost-Aware Routing

Automatically routes to the lowest cost model based on cost_per_1k_tokens. Ideal for cost optimization.

configs = [
    ModelConfig(cost_per_1k_tokens=0.03),   # GPT-4 - $0.03/1k tokens
    ModelConfig(cost_per_1k_tokens=0.002),  # GPT-3.5 - $0.002/1k tokens
    ModelConfig(cost_per_1k_tokens=0.015),  # Claude - $0.015/1k tokens
]

manager = MultiModelManager(
    models=models,
    model_configs=configs,
    strategy=RoutingStrategy.COST_AWARE
)

Round-Robin Routing

Distributes requests evenly across all available models. Great for load balancing.

manager = MultiModelManager(
    models=models,
    strategy=RoutingStrategy.ROUND_ROBIN
)

Least-Used Routing

Routes to the model with the fewest total requests. Helps balance usage across models.

manager = MultiModelManager(
    models=models,
    strategy=RoutingStrategy.LEAST_USED
)

Custom Strategies

You can provide a custom routing function for advanced use cases:

def custom_strategy(models, configs, usage_stats, available_models):
    """Custom strategy: prefer models with highest success rate."""
    best_model = available_models[0]
    best_rate = 0.0
    
    for idx in available_models:
        stats = usage_stats.get(idx)
        if stats and stats.total_requests > 0:
            success_rate = stats.successful_requests / stats.total_requests
            if success_rate > best_rate:
                best_rate = success_rate
                best_model = idx
    
    return best_model

manager = MultiModelManager(
    models=models,
    strategy=custom_strategy
)

Structured Output

Get Pydantic-validated responses from any model, with automatic fallback for models without native structured output support.

from pydantic import BaseModel, Field

class Person(BaseModel):
    """Information about a person."""
    name: str = Field(description="The person's full name")
    age: int = Field(description="The person's age in years")
    occupation: str = Field(description="The person's job or profession")

# Create structured output runnable
structured_manager = manager.with_structured_output(Person)

# Get validated Pydantic object
person = structured_manager.invoke("Tell me about Albert Einstein")
print(f"{person.name} was {person.age} years old and worked as a {person.occupation}")
# Output: Albert Einstein was 76 years old and worked as a Theoretical Physicist

The structured output handler automatically:

• Detects if the model has native structured output support
• Uses native support when available for better performance
• Falls back to prompt injection and JSON parsing when needed
• Validates all responses against your Pydantic schema

Rate Limiting and Fallback

Configure per-model rate limits and automatic fallback behavior:

from langchain_fused_model import ModelConfig

configs = [
    ModelConfig(
        priority=100,
        max_rpm=60,       # 60 requests per minute
        max_rps=2,        # 2 requests per second
        timeout=30.0,     # 30 second timeout
        retry_on_errors=[TimeoutError, ConnectionError]
    ),
    ModelConfig(
        priority=50,
        max_rpm=120,      # Fallback model with higher limits
    ),
]

manager = MultiModelManager(
    models=models,
    model_configs=configs,
    strategy=RoutingStrategy.PRIORITY,
    default_fallback=True  # Enable automatic fallback
)

# Automatically falls back if rate limit exceeded or errors occur
response = manager.invoke("Your prompt here")

When a model fails or hits rate limits:

1. The manager automatically selects the next available model
2. A cooldown period is set for rate-limited models
3. The request is retried with the new model
4. All failures are logged for monitoring

LangChain Integration

The MultiModelManager works seamlessly with all LangChain features:

Chains

from langchain.prompts import ChatPromptTemplate
from langchain.schema.output_parser import StrOutputParser

# Use in chains with the pipe operator
prompt = ChatPromptTemplate.from_template("Tell me a joke about {topic}")
chain = prompt | manager | StrOutputParser()

result = chain.invoke({"topic": "programming"})
print(result)

Batch Processing

# Process multiple inputs in parallel
questions = [
    "What is Python?",
    "What is JavaScript?",
    "What is Rust?"
]

responses = manager.batch(questions)
for response in responses:
    print(response.content)

Streaming (if supported by underlying models)

# Stream responses token by token
for chunk in manager.stream("Write a long story about AI"):
    print(chunk.content, end="", flush=True)

Agents

from langchain.agents import AgentExecutor, create_openai_functions_agent
from langchain.tools import Tool

# Use as the LLM for agents
agent = create_openai_functions_agent(manager, tools, prompt)
agent_executor = AgentExecutor(agent=agent, tools=tools)

result = agent_executor.invoke({"input": "What's the weather in Paris?"})

Usage Statistics

Monitor model performance and usage:

# Get statistics for all models
stats = manager._usage_tracker.get_all_stats()

for model_idx, stat in stats.items():
    print(f"\nModel {model_idx} ({models[model_idx]._llm_type}):")
    print(f"  Total requests: {stat.total_requests}")
    print(f"  Successful: {stat.successful_requests}")
    print(f"  Failed: {stat.failed_requests}")
    
    if stat.total_requests > 0:
        success_rate = stat.successful_requests / stat.total_requests
        print(f"  Success rate: {success_rate:.2%}")
    
    print(f"  Total tokens: {stat.total_tokens}")
    
    if stat.last_used:
        import datetime
        last_used = datetime.datetime.fromtimestamp(stat.last_used)
        print(f"  Last used: {last_used}")

# Get statistics for a specific model
model_0_stats = manager._usage_tracker.get_stats(0)
print(f"Model 0 has handled {model_0_stats.total_requests} requests")

Advanced Configuration

Complete Configuration Example

from langchain_fused_model import MultiModelManager, ModelConfig, RoutingStrategy

configs = [
    ModelConfig(
        priority=100,              # Highest priority
        max_rpm=60,                # Rate limits
        max_rps=2,
        cost_per_1k_tokens=0.03,   # Cost tracking
        timeout=30.0,              # Request timeout
        retry_on_errors=[          # Custom retry conditions
            TimeoutError,
            ConnectionError,
        ]
    ),
    ModelConfig(
        priority=50,
        max_rpm=120,
        max_rps=5,
        cost_per_1k_tokens=0.002,
        timeout=20.0,
    ),
]

manager = MultiModelManager(
    models=models,
    model_configs=configs,
    strategy=RoutingStrategy.PRIORITY,
    default_fallback=True
)

Logging Configuration

The package uses Python's standard logging module:

import logging

# Enable debug logging to see model selection decisions
logging.basicConfig(level=logging.DEBUG)
logger = logging.getLogger('langchain_fused_model')
logger.setLevel(logging.DEBUG)

# Now you'll see detailed logs about model selection and fallback
response = manager.invoke("Test prompt")

Examples

Check out the examples/ directory for Jupyter notebooks demonstrating:

• basic_usage.ipynb: Getting started with MultiModelManager
• routing_strategies.ipynb: Comparing all routing strategies
• structured_output.ipynb: Working with Pydantic models and structured data

Requirements

• Python 3.9+
• langchain-core >= 0.1.0
• pydantic >= 2.0.0

Optional dependencies for specific providers:

• langchain-openai (for OpenAI models)
• langchain-anthropic (for Anthropic models)
• langchain-google-genai (for Google models)

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

To set up for development:

git clone https://github.com/yourusername/langchain-fused-model
cd langchain-fused-model
pip install -e ".[dev]"
pytest tests/

License

MIT License - see LICENSE file for details.

Support

• GitHub Issues: https://github.com/yourusername/langchain-fused-model/issues
• Documentation: https://github.com/yourusername/langchain-fused-model#readme
• Examples: See the examples/ directory for Jupyter notebooks

---

Note: This package is designed to work with any LangChain-compatible ChatModel. Make sure to install the appropriate provider packages (e.g., langchain-openai, langchain-anthropic) for the models you want to use.