langgraph/how-tos/create-react-agent-structured-output/

[giscus[bot]]

langgraph/how-tos/create-react-agent-structured-output/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/how-tos/create-react-agent-structured-output/

[Intraraksa]

I got error
create_react_agent() got an unexpected keyword argument 'response_format'

[jma-hdz]

maybe you need to check your langgraph version (pip list, etc), for example, you could try with the 0.2.71 version

[sanggi-wjg]

use this path

from langgraph.prebuilt import create_react_agent

[hinthornw]

Try upgrading your langgraph-prebiult package

[edgarlcs]

For some reason the agent responds well in the message responses but returns something completely different on the structured output every time. Has anyone encountered this issue?

[limcolin]

Is there a JS version to achieve this using 2 LLMs?

Binding output as tool is giving me invalid_tool_calls when streaming with streamEvents

[kapis]

You guys aware you can have both tools and structured output with OpenAI models? No need to make is a tool.

[hinthornw]

Yes. That is why response_format uses structured output instead of tools.

[kapis]

Then why "Prebuilt ReAct makes an additional LLM call at the end of the ReAct loop to produce a structured output response."? With OpenAI, every LLM call can return both tool calls and structured output. So no additional LLM call should be required.

[dougzec]

@kapis is there a langgraph implementation that allows leveraging these OpenAI API features?

[sree24lakshmi]

Can anyone please suggest if there is a way to overcome context window length when working with SQLDatabasetoolkit. The agent runs into error with stop reason=length as the context length exceeding in case of large output returned by database. Appreciate your help thank you

[Khalid-Elejla]

You can modify the sql query tool within SQLDatabasetoolkit (i think there is 3 or 4 other tools) and make it instead of returning the result saving it to a dataframe, csv or txt file and return just some metadata like dataframe id (to retreive from memory) or file path, name, number of columns maybe also a header
You can use also content and artifact when you redefine your tool

[edmald]

Hi Guys,

I'm working with this prebuilt agent:
agent_executor = create_react_agent(
llm,
tools=tools,
prompt=system_prompt,
checkpointer=memory,
debug=False
)

Configuration for the agent executor

config_1 = {"configurable": {"thread_id": "1"}}

One of the tools I'm using is defined like this:

def tool_plot_field_production_rates_wrapper(args: str) -> str:
"""
Generates field production rate plots for all scenarios (P10, P50, P90), showing oil, water, and liquid production
on the primary axis and gas production on the secondary axis.

The tool supports customization of line colors (default: oil in green, gas in red, water in blue, liquid in black).
Use this tool when the query involves comparing field-wide production across different scenarios.
The output includes both interactive plots and downloadable static images for each scenario.
"""
try:
    # Extract parameters from the query
    params = extract_field_production_params(args)
    
    # Call the plotting function with extracted parameters
    results = plot_field_production_rates(
        oil_color=params["oil_color"],
        gas_color=params["gas_color"],
        water_color=params["water_color"],
        liquid_color=params["liquid_color"]
    )
    ...

Tool registration is done as follows:

tools = [
Tool(
name="plot_field_production_rates",
func=tool_plot_field_production_rates_wrapper,
description=(
"Generates comprehensive field-level production charts for ALL scenarios (P10, P50, P90) in a single run. "
"Displays oil, water, and liquid on the primary axis, and gas on the secondary axis, for the ENTIRE FIELD. "
"Use this tool specifically when the query involves field-wide comparisons or mentions multiple scenarios. "
"Handles all color customization (e.g., 'use green for oil') in one execution—no need to call it multiple times. "
"Keywords like 'by scenario' or 'for all scenarios' are good indicators this is the right tool."
)
)
]
Issue I'm seeing:
When a user makes a query, the ReAct agent tries to interpret it before calling the tool. However, the args parameter received in the tool is not the original user question—it's already been altered or reformatted.

Question:
Why is the args input changed from the original user question? Is this expected behavior from the ReAct agent framework? I want to understand at what point and why the user input is being transformed before it's passed to the tool.

Any possibility to stop the react agent doing this?

Any help is good!!!

[AkashKandra]

I am working on structured data retrieval using this create_react_agent. For the first user input, it is generating the response. If the second user input is like, regenerate the sql query, it has to call the generate_sql_query tool and regenrate the sql query.
Now I want to know how can I pass the message history while asking the second question.

[velvinnn]

Has any one tried using local huggingface models? I tried multiple LLMs with the huggingface pipeline, but none of them successfully called the tools.

Please see my code below:

from langchain_huggingface import HuggingFacePipeline

llm = HuggingFacePipeline.from_model_id(
    model_id=model_name,
    task="text-generation",
    device_map='auto',
    pipeline_kwargs=dict(
        max_new_tokens=512,
        # do_sample=False,
        temperature=0.2,
        repetition_penalty=1.03,
    ),
)

model = ChatHuggingFace(llm=llm)

[velvinnn]

the model_name is the huggingface id, such as 'meta-llama/Llama-3.3-70B-Instruct', 'mistralai/Mistral-7B-Instruct-v0.3', Groq/Llama-3-Groq-8B-Tool-Use', 'meta-llama/Llama-4-Maverick-17B-128E-Instruct','google/gemma-3-27b-it'

[haithamafifi]

I probably know what you are talking about. You need to invest time in

• prompt Tuning.
• better tool description

at least that was my problem, 7/8 trials the tool calling work with opensource models. Invest a bit in the prompting and description and you may get better results

[baglan1]

In JS version, you can use ChatOpenAI.bind to add both structured response and a tool call. The following code shows how create an LLM that calls a tool and binds a reponse form at to the LLM:

const structuremLLM = model.bindTools(tools, { strict: true })
      .bind({
        response_format: {
          type: "json_schema",
          json_schema: {
            name: "resource",
            description: "Resource for the lesson",
            schema: zodSchema, // Use the Zod object directly
          },
        },
      });

More parameters can be set to the OpenAI LLM in langchain. Here's a list of parameters that can be set including modalities and reasoning.