Replace TypedDict with Pydantic BaseModel

[SandaruRF]

Switched TypedDict to Pydantic BaseModel for structured output.
This improves validation and better type enforcement.
TypeDict gives errors when using Gemini-2.0-flash model.
This is compatible with Anthropic, OpenAI, and Gemini models.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langgraph-docs-preview	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Mar 21, 2025 2:38pm

[hinthornw]

I don't think we want to make this change. People use differnet versions of pydantic and run into random compatibility issues, so I feel more comfortable using it less.

[vbarda]

agreed, i don't think it's necessary

[vbarda]

@hinthornw actually i think it should be fine, will double check that this doesn't cause any issues

[SandaruRF]

TypedDict gives some warnings and at the end resource exhausted when using with Gemini.

ResourceExhausted: 429 Resource has been exhausted (e.g. check quota).
During task with name 'agent' and id 'bdae53f5-2b37-33ad-9fa8-e6f37e8076b7'
During task with name 'coder' and id '92669bdd-2a7b-0320-2746-d66f998a4eae'

However when using Pydantic BaseModel, there are no warnings and resource exhaustion doesn't occur. And this works fine with other models.

[eyurtsev]

https://developer.mozilla.org/en-US/docs/Web/HTTP/Reference/Status/429 @SandaruRF

You should try changing back to back. It's unlikely that the 429 is related to typeddict vs. pydantic model unless it's formulating an incorrect payload and failing w/ some 4xx (and we're automatically retrying in the tool node @vbarda )

@SandaruRF could you confirm that you can repeat the failure by changing from typeddict to pydantic and back?

[SandaruRF]

@eyurtsev I ran tests comparing TypedDict and Pydantic BaseModel. When using TypedDict I get a resource exhausted error on the first run. But when I use Pydantic it runs successfully. I think there is a possible issue with request formatting, retries or payload handling. This happens when using Gemini.

[ccurme]

@eyurtsev I ran tests comparing TypedDict and Pydantic BaseModel. When using TypedDict I get a resource exhausted error on the first run. But when I use Pydantic it runs successfully. I think there is a possible issue with request formatting, retries or payload handling. This happens when using Gemini.

@SandaruRF can you please share some details about the model you are instantiating? e.g., if you are using langchain-google-genai or langchain-google-vertexai, the model string, the versions of relevant langchain packages, etc.

[SandaruRF]

@eyurtsev I ran tests comparing TypedDict and Pydantic BaseModel. When using TypedDict I get a resource exhausted error on the first run. But when I use Pydantic it runs successfully. I think there is a possible issue with request formatting, retries or payload handling. This happens when using Gemini.

@SandaruRF can you please share some details about the model you are instantiating? e.g., if you are using langchain-google-genai or langchain-google-vertexai, the model string, the versions of relevant langchain packages, etc.

@ccurme I am using langchain-google-genai with gemini-2.0-flash model. The versions of langchain packages are,
langgraph - 0.3.21
langchain-community - 0.3.20
langchain-google-genai - 2.1.2
langchain-experimental - 0.3.4

[ccurme]

@SandaruRF could you open an issue with a minimal reproducible example? I'm unable to reproduce using latest langchain-google-genai (which you are using). Here is my attempt:

from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.prebuilt import create_react_agent
from typing_extensions import TypedDict


llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash-001")


def get_weather(location: str) -> str:
    """Get weather at a location."""
    return "It's sunny."


class ResponseFormat(TypedDict):
    """Respond to the user in this format."""
    my_special_output: str


graph = create_react_agent(
    llm,
    tools=[get_weather],
    response_format=ResponseFormat
)

state = graph.invoke({"messages": "What's the weather in Boston?"})
state["structured_response"]  # {'my_special_output': "It's sunny in Boston"}

[SandaruRF]

@SandaruRF could you open an issue with a minimal reproducible example? I'm unable to reproduce using latest langchain-google-genai (which you are using). Here is my attempt:

from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.prebuilt import create_react_agent
from typing_extensions import TypedDict


llm = ChatGoogleGenerativeAI(model="gemini-2.0-flash-001")


def get_weather(location: str) -> str:
    """Get weather at a location."""
    return "It's sunny."


class ResponseFormat(TypedDict):
    """Respond to the user in this format."""
    my_special_output: str


graph = create_react_agent(
    llm,
    tools=[get_weather],
    response_format=ResponseFormat
)

state = graph.invoke({"messages": "What's the weather in Boston?"})
state["structured_response"]  # {'my_special_output': "It's sunny in Boston"}

Sure