[Bug]: Pydantic model's description is not taken into account when generating structured outputs

[venki-lfc]

Your current environment

vllm == 0.11.0
langchain==1.2.0
langchain-core== 1.2.6
langchain-openai == 1.1.6

🐛 Describe the bug

When using with_structured_output from langchain's ChatOpenAI, the model response does not seem to follow the rules mentioned in description field. Here is an example:

# Define a simple Pydantic schema
class CarDescription(BaseModel):
    origin: str = Field(description="Country of origin in Uppercase")
    model: str = Field(description="Model of the car in lowercase")
    car_type: Literal["Sports", "Luxury", "Economic"] = Field(description="Type of the car")


# Wrap with structured output
structured_llm = llm.with_structured_output(CarDescription)

# Generate Result
result = structured_llm.invoke("Create a car description for VW")

The results looks like this:
origin='Germany' model='Golf GTI' car_type='Sports'

It can be seen that the LLM failed to generate the origin and model in the respective cases as mentioned in their descriptions. I served Qwen3-30B-A3B-Instruct using vllm to get the result.

I tested this with GPT-5-mini and the results were in the correct cases. So I assume vLLM somehow does not read the description of each key but just the name of the key. Can you please fix this issue?

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[sciafri]

Also encountered this issue

It appears pydantic.Field is not being properly provided to the LLM. For example:

PHONE_NUMBER_REGEX = r"^(\+\d{1,2}\s)?\(?\d{3}\)?[\s.-]\d{3}[\s.-]\d{4}$"


class ContactInfo(BaseModel):
    name: str = Field(description="The name of the person")
    email: str = Field(description="The email address of the person. If none is provided, do not make one up, instead fill in with x@x.x or None type")
    phone: str = Field(pattern=PHONE_NUMBER_REGEX, description="The phone number of the person. In the format if no area code given put XXX as the area code. If no country code is provided, provide +54 this is the only time you are allowed to make something up or assume.")


content = "John Smith john@smith.com 888-123-4312"
completion = client.chat.completions.parse(
        model=model,
        seed=1234,
        temperature=.5,
        messages=[
            {
                "role": "system",
                "content": "extract information provided below following the word content, use json schema to extract information. only look at information which follows"
            },
            {
                "role": "user",
                "content": f"{content}"
            },
        ],
        response_format=ContactInfo
  )

response = completion.choices[0].message.content
print(response)

Model gpt-4o-2024-08-06 from openai provides the following output (note: this is true for both client.beta.chat.completions.parse and client.chat.completions.parse)

'{"name":"John Smith","email":"john@smith.com","phone":"+54 888-123-4312"}'

vllm provides the following:

completion = client_vllm.chat.completions.parse(
        model=model,
        seed=1234,
        temperature=.5,
        messages=[
            {
                "role": "system",
                "content": "extract information provided below following the word content, use json schema to extract information. only look at information which follows"
            },
            {
                "role": "user",
                "content": f"{content}"
            },
        ],
        response_format=ContactInfo
  )
response = completion.choices[0].message.content
print(response)

'{\n "name": "John Smith",\n "email": "john@smith.com",\n "phone": "888-123-4312"\n}'

This means that these fields are being completely ignored as no +54 was added. I wasn't able to get it to follow any instructions in the description such as provide 555-555-5555 if no number is provided etc. Using model llama 3.2 8B.

My code above is using directly openai api and pydantic and was able to get the response to differ when switching to openai's endpoints, so very likely the bug exists within vLLM.