langchain-dev-utils 1.2.2

A practical utility library for LangChain and LangGraph development

🦜️🧰 langchain-dev-utils

A utility library for LangChain and LangGraph development.

📚 English • 中文

This is the English version. For the Chinese version, please visit Chinese Documentation

langchain-dev-utils is a utility library focused on enhancing the development experience with LangChain and LangGraph. It provides a series of out-of-the-box utility functions that can both reduce repetitive code writing and improve code consistency and readability. By simplifying development workflows, this library helps you prototype faster, iterate more smoothly, and create clearer, more reliable LLM-based AI applications.

🚀 Installation

pip install -U langchain-dev-utils

# Install the full-featured version:
pip install -U langchain-dev-utils[standard]

📦 Core Features

1. Model Management

In langchain, the init_chat_model/init_embeddings functions can be used to initialize chat model instances/embedding model instances, but the model providers they support are relatively limited. This module provides a registration function (register_model_provider/register_embeddings_provider) to register any model provider for subsequent model loading using load_chat_model / load_embeddings.

1.1 Chat Model Management

Mainly consists of the following two functions:

• register_model_provider: Register a chat model provider
• load_chat_model: Load a chat model

register_model_provider parameter description:

• provider_name: Model provider name, used as an identifier for subsequent model loading
• chat_model: Chat model, can be a ChatModel or a string (currently supports "openai-compatible")
• base_url: The API address of the model provider (optional, valid for both types of chat_model, but mainly used when chat_model is a string and is "openai-compatible")
• provider_config: Relevant configuration for the model provider (optional, valid when chat_model is a string and is "openai-compatible"), can configure some provider-related parameters, such as whether to support structured output in json_mode, list of supported tool_choices, etc.

load_chat_model parameter description:

• model: Chat model name, type str
• model_provider: Chat model provider name, type str, optional
• kwargs: Additional parameters passed to the chat model class, e.g., temperature, top_p, etc.

Example for integrating a qwen3-4b model deployed using vllm:

from langchain_dev_utils.chat_models import (
    register_model_provider,
    load_chat_model,
)

# Register model provider
register_model_provider(
    provider_name="vllm",
    chat_model="openai-compatible",
    base_url="http://localhost:8000/v1",
)

# Load model
model = load_chat_model("vllm:qwen3-4b")
print(model.invoke("Hello"))

1.2 Embedding Model Management

Mainly consists of the following two functions:

• register_embeddings_provider: Register an embedding model provider
• load_embeddings: Load an embedding model

register_embeddings_provider parameter description:

• provider_name: Embedding model provider name, used as an identifier for subsequent model loading
• embeddings_model: Embedding model, can be Embeddings or a string (currently supports "openai-compatible")
• base_url: The API address of the Embedding model provider (optional, valid for both types of embeddings_model, but mainly used when embeddings_model is a string and is "openai-compatible")

load_embeddings parameter description:

• model: Embedding model name, type str
• provider: Embedding model provider name, type str, optional
• kwargs: Other additional parameters

Example for integrating a qwen3-embedding-4b model deployed using vllm:

from langchain_dev_utils.embeddings import register_embeddings_provider, load_embeddings

# Register embedding model provider
register_embeddings_provider(
    provider_name="vllm",
    embeddings_model="openai-compatible",
    base_url="http://localhost:8000/v1",
)

# Load embedding model
embeddings = load_embeddings("vllm:qwen3-embedding-4b")
emb = embeddings.embed_query("Hello")
print(emb)

For more information about model management, please refer to: Chat Model Management, Embedding Model Management

2. Message Conversion

Includes the following features:

• Merge reasoning content into the final response
• Stream content merging
• Content formatting tools

2.1 Stream Content Merging

For stream responses obtained using stream() and astream(), you can use merge_ai_message_chunk to merge them into a final AIMessage.

merge_ai_message_chunk parameter description:

• chunks: List of AIMessageChunk

chunks = list(model.stream("Hello"))
merged = merge_ai_message_chunk(chunks)

2.2 Format List Content

For a list, you can use format_sequence to format it.

format_sequence parameter description:

• inputs: A list containing any of the following types:
  • langchain_core.messages: HumanMessage, AIMessage, SystemMessage, ToolMessage
  • langchain_core.documents.Document
  • str
• separator: String used to join the content, defaults to "-".
• with_num: If True, add a numeric prefix to each item (e.g., "1. Hello"), defaults to False.

text = format_sequence([
    "str1",
    "str2",
    "str3"
], separator="\n", with_num=True)

For more information about message conversion, please refer to: Message Processing, Format List Content

3. Tool Calling

Includes the following features:

• Check and parse tool calls
• Add human-in-the-loop functionality

3.1 Check and Parse Tool Calls

has_tool_calling and parse_tool_calling are used to check and parse tool calls.

has_tool_calling parameter description:

• message: AIMessage object

parse_tool_calling parameter description:

• message: AIMessage object
• first_tool_call_only: Whether to only check the first tool call

import datetime
from langchain_core.tools import tool
from langchain_dev_utils.tool_calling import has_tool_calling, parse_tool_calling

@tool
def get_current_time() -> str:
    """Get the current timestamp"""
    return str(datetime.datetime.now().timestamp())

response = model.bind_tools([get_current_time]).invoke("What time is it?")

if has_tool_calling(response):
    name, args = parse_tool_calling(
        response, first_tool_call_only=True
    )
    print(name, args)

3.2 Add Human-in-the-Loop Functionality

• human_in_the_loop: For synchronous tool functions
• human_in_the_loop_async: For asynchronous tool functions

Both can accept a handler parameter for custom breakpoint return and response handling logic.

from langchain_dev_utils import human_in_the_loop
from langchain_core.tools import tool
import datetime

@human_in_the_loop
@tool
def get_current_time() -> str:
    """Get the current timestamp"""
    return str(datetime.datetime.now().timestamp())

For more information about tool calling, please refer to: Add Human-in-the-Loop Support, Tool Call Processing

4. Agent Development

Includes the following features:

• Predefined agent factory functions
• Common middleware components

4.1 Agent Factory Functions

In LangChain v1, the officially provided create_agent function can be used to create a single agent, where the model parameter supports passing a BaseChatModel instance or a specific string (when passing a string, it is limited to the models supported by init_chat_model). To extend the flexibility of specifying models via strings, this library provides a functionally identical create_agent function, allowing you to directly use models supported by load_chat_model (requires prior registration).

Usage example:

from langchain_dev_utils.agents import create_agent
from langchain.agents import AgentState

agent = create_agent("vllm:qwen3-4b", tools=[get_current_time], name="time-agent")
response = agent.invoke({"messages": [{"role": "user", "content": "What time is it?"}]})
print(response)

4.2 Middleware

Provides some commonly used middleware components. Below are examples of SummarizationMiddleware and PlanMiddleware.

SummarizationMiddleware is used for agent summarization.

PlanMiddleware is used for agent planning.

from langchain_dev_utils.agents.middleware import (
    SummarizationMiddleware,
    PlanMiddleware,
)

agent=create_agent(
    "vllm:qwen3-4b",
    name="plan-agent",
    middleware=[PlanMiddleware(), SummarizationMiddleware(model="vllm:qwen3-4b")]
)
response = agent.invoke({"messages": [{"role": "user", "content": "Give me a travel plan to New York"}]}))
print(response)

For more information about agent development and all built-in middleware, please refer to: Prebuilt Agent Functions, Middleware

5. State Graph Orchestration

Includes the following features:

• Sequential graph orchestration
• Parallel graph orchestration

5.1 Sequential Graph Orchestration

Sequential graph orchestration: Uses create_sequential_pipeline, supported parameters:

• sub_graphs: List of state graphs to combine (must be StateGraph instances)
• state_schema: State Schema for the final generated graph
• graph_name: Name of the final generated graph (optional)
• context_schema: Context Schema for the final generated graph (optional)
• input_schema: Input Schema for the final generated graph (optional)
• output_schema: Output Schema for the final generated graph (optional)
• checkpoint: LangGraph persistence Checkpoint (optional)
• store: LangGraph persistence Store (optional)
• cache: LangGraph Cache (optional)

from langchain.agents import AgentState
from langchain_core.messages import HumanMessage
from langchain_dev_utils.agents import create_agent
from langchain_dev_utils.pipeline import create_sequential_pipeline
from langchain_dev_utils.chat_models import register_model_provider

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

# Build sequential pipeline (all sub-graphs execute sequentially)
graph = create_sequential_pipeline(
    sub_graphs=[
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_time],
            system_prompt="You are a time query assistant, can only answer the current time. If the question is unrelated to time, please directly answer that you cannot answer.",
            name="time_agent",
        ),
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_weather],
            system_prompt="You are a weather query assistant, can only answer the current weather. If the question is unrelated to weather, please directly answer that you cannot answer.",
            name="weather_agent",
        ),
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_user],
            system_prompt="You are a user query assistant, can only answer the current user. If the question is unrelated to user, please directly answer that you cannot answer.",
            name="user_agent",
        ),
    ],
    state_schema=AgentState,
)

response = graph.invoke({"messages": [HumanMessage("Hello")]})
print(response)

5.2 Parallel Graph Orchestration

Parallel graph orchestration: Uses create_parallel_pipeline, supported parameters:

• sub_graphs: List of state graphs to combine
• state_schema: State Schema for the final generated graph
• branches_fn: Parallel branch function, returns a list of Send objects to control parallel execution
• graph_name: Name of the final generated graph (optional)
• context_schema: Context Schema for the final generated graph (optional)
• input_schema: Input Schema for the final generated graph (optional)
• output_schema: Output Schema for the final generated graph (optional)
• checkpoint: LangGraph persistence Checkpoint (optional)
• store: LangGraph persistence Store (optional)
• cache: LangGraph Cache (optional)

from langchain_dev_utils.pipeline import create_parallel_pipeline

# Build parallel pipeline (all sub-graphs execute in parallel)
graph = create_parallel_pipeline(
    sub_graphs=[
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_time],
            system_prompt="You are a time query assistant, can only answer the current time. If the question is unrelated to time, please directly answer that you cannot answer.",
            name="time_agent",
        ),
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_weather],
            system_prompt="You are a weather query assistant, can only answer the current weather. If the question is unrelated to weather, please directly answer that you cannot answer.",
            name="weather_agent",
        ),
        create_agent(
            model="vllm:qwen3-4b",
            tools=[get_current_user],
            system_prompt="You are a user query assistant, can only answer the current user. If the question is unrelated to user, please directly answer that you cannot answer.",
            name="user_agent",
        ),
    ],
    state_schema=AgentState,
)
response = graph.invoke({"messages": [HumanMessage("Hello")]})
print(response)

For more information about state graph orchestration, please refer to: State Graph Orchestration Pipeline

💬 Join the Community

• GitHub Repository — Browse source code, submit Pull Requests
• Issue Tracker — Report bugs or suggest improvements
• We welcome contributions in all forms — whether code, documentation, or usage examples. Let's build a more powerful and practical LangChain development ecosystem together!