Overview

Middleware components are pluggable modules specifically designed for LangChain's pre-built Agents, aiming to achieve fine-grained control over the Agent's internal behavior. In addition to the built-in middleware provided by the LangChain framework, this library further enriches middleware support based on practical application scenarios.

The middleware provided by this library includes:

• PlanMiddleware: Task planning, breaking down complex tasks into ordered subtasks
• ModelRouterMiddleware: Dynamically routing input to the most suitable model based on content
• HandoffAgentMiddleware: Flexibly handing off tasks between multiple sub-agents
• ToolCallRepairMiddleware: Automatically repairing invalid tool calls from LLMs
• FormatPromptMiddleware: Dynamically formatting placeholders in system prompts

Furthermore, this library extends the functionality of official middleware, enhancing model configuration usability by supporting model specification through string parameters:

• SummarizationMiddleware
• LLMToolSelectorMiddleware
• ModelFallbackMiddleware
• LLMToolEmulator

Note

In the subsequent examples, we import the create_agent function from langchain_dev_utils.agents rather than langchain.agents. This is because this library provides a function that is functionally identical to the official create_agent, but has been extended to support specifying models via strings. This allows for the direct use of models registered via register_model_provider without the need to initialize and pass in a model instance.

Before running the examples, please ensure:

1.Register the vllm model provider

register_model_provider(
    "vllm",
    "openai-compatible",
    base_url="http://localhost:8000/v1",
)

2.Import the create_agent function from langchain_dev_utils.agents

from langchain_dev_utils.agents import create_agent