chat ollama with_structured_output fails with union type

[eyurtsev]

Privileged issue

• [X] I am a LangChain maintainer, or was asked directly by a LangChain maintainer to create an issue here.

Issue Content

from typing import Union
from langchain_ollama import ChatOllama

llm = ChatOllama(model="llama3.1", temperature=0)

# Pydantic
class Joke(BaseModel):
    """Joke to tell user."""
    setup: str = Field(description="The setup of the joke")
    punchline: str = Field(description="The punchline to the joke")
    rating: Optional[int] = Field(
        default=None, description="How funny the joke is, from 1 to 10"
    )

class ConversationalResponse(BaseModel):
    """Respond in a conversational manner. Be kind and helpful."""
    response: str = Field(description="A conversational response to the user's query")

class FinalResponse(BaseModel):
    final_output: Union[Joke, ConversationalResponse]

structured_llm = llm.with_structured_output(FinalResponse)
structured_llm.invoke("Tell me a joke about cats")

raise here: https://github.com/langchain-ai/langchain/pull/27415

[eyurtsev]

stack trace:

---------------------------------------------------------------------------
ValidationError                           Traceback (most recent call last)
Cell In[31], line 25
     22     final_output: Union[Joke, ConversationalResponse]
     24 structured_llm = llm.with_structured_output(FinalResponse)
---> 25 structured_llm.invoke("Tell me a joke about cats")

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/base.py:3024](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/base.py#line=3023), in RunnableSequence.invoke(self, input, config, **kwargs)
   3022             input = context.run(step.invoke, input, config, **kwargs)
   3023         else:
-> 3024             input = context.run(step.invoke, input, config)
   3025 # finish the root run
   3026 except BaseException as e:

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/base.py:193](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/base.py#line=192), in BaseOutputParser.invoke(self, input, config, **kwargs)
    186 def invoke(
    187     self,
    188     input: Union[str, BaseMessage],
    189     config: Optional[RunnableConfig] = None,
    190     **kwargs: Any,
    191 ) -> T:
    192     if isinstance(input, BaseMessage):
--> 193         return self._call_with_config(
    194             lambda inner_input: self.parse_result(
    195                 [ChatGeneration(message=inner_input)]
    196             ),
    197             input,
    198             config,
    199             run_type="parser",
    200         )
    201     else:
    202         return self._call_with_config(
    203             lambda inner_input: self.parse_result([Generation(text=inner_input)]),
    204             input,
    205             config,
    206             run_type="parser",
    207         )

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/base.py:1927](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/base.py#line=1926), in Runnable._call_with_config(self, func, input, config, run_type, serialized, **kwargs)
   1923     context = copy_context()
   1924     context.run(_set_config_context, child_config)
   1925     output = cast(
   1926         Output,
-> 1927         context.run(
   1928             call_func_with_variable_args,  # type: ignore[arg-type]
   1929             func,  # type: ignore[arg-type]
   1930             input,  # type: ignore[arg-type]
   1931             config,
   1932             run_manager,
   1933             **kwargs,
   1934         ),
   1935     )
   1936 except BaseException as e:
   1937     run_manager.on_chain_error(e)

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/config.py:396](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/runnables/config.py#line=395), in call_func_with_variable_args(func, input, config, run_manager, **kwargs)
    394 if run_manager is not None and accepts_run_manager(func):
    395     kwargs["run_manager"] = run_manager
--> 396 return func(input, **kwargs)

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/base.py:194](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/base.py#line=193), in BaseOutputParser.invoke.<locals>.<lambda>(inner_input)
    186 def invoke(
    187     self,
    188     input: Union[str, BaseMessage],
    189     config: Optional[RunnableConfig] = None,
    190     **kwargs: Any,
    191 ) -> T:
    192     if isinstance(input, BaseMessage):
    193         return self._call_with_config(
--> 194             lambda inner_input: self.parse_result(
    195                 [ChatGeneration(message=inner_input)]
    196             ),
    197             input,
    198             config,
    199             run_type="parser",
    200         )
    201     else:
    202         return self._call_with_config(
    203             lambda inner_input: self.parse_result([Generation(text=inner_input)]),
    204             input,
    205             config,
    206             run_type="parser",
    207         )

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/openai_tools.py:298](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/openai_tools.py#line=297), in PydanticToolsParser.parse_result(self, result, partial)
    296             continue
    297         else:
--> 298             raise e
    299 if self.first_tool_only:
    300     return pydantic_objects[0] if pydantic_objects else None

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/openai_tools.py:293](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/langchain_core/output_parsers/openai_tools.py#line=292), in PydanticToolsParser.parse_result(self, result, partial)
    288         msg = (
    289             f"Tool arguments must be specified as a dict, received: "
    290             f"{res['args']}"
    291         )
    292         raise ValueError(msg)
--> 293     pydantic_objects.append(name_dict[res["type"]](**res["args"]))
    294 except (ValidationError, ValueError) as e:
    295     if partial:

File [~/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/pydantic/main.py:212](http://localhost:8888/home/eugene/.pyenv/versions/3.11.4/envs/core_3_11_4/lib/python3.11/site-packages/pydantic/main.py#line=211), in BaseModel.__init__(self, **data)
    210 # `__tracebackhide__` tells pytest and some other tools to omit this function from tracebacks
    211 __tracebackhide__ = True
--> 212 validated_self = self.__pydantic_validator__.validate_python(data, self_instance=self)
    213 if self is not validated_self:
    214     warnings.warn(
    215         'A custom validator is returning a value other than `self`.\n'
    216         "Returning anything other than `self` from a top level model validator isn't supported when validating via `__init__`.\n"
    217         'See the `model_validator` docs (https://docs.pydantic.dev/latest/concepts/validators/#model-validators) for more details.',
    218         category=None,
    219     )

ValidationError: 2 validation errors for FinalResponse
final_output.Joke
  Input should be a valid dictionary or instance of Joke [type=model_type, input_value='Why did the cat join a b...be the purr-cussionist!', input_type=str]
    For further information visit https://errors.pydantic.dev/2.9/v/model_type
final_output.ConversationalResponse
  Input should be a valid dictionary or instance of ConversationalResponse [type=model_type, input_value='Why did the cat join a b...be the purr-cussionist!', input_type=str]
    For further information visit https://errors.pydantic.dev/2.9/v/model_type

[eyurtsev]

This is likely not an issue with the implementation, but the model failing to actually output appropriate structured_output -- so performance issue w/ the underlying model.

The TODO item is to figure out how to improve error messages so such issues are easier for the user to understand

[ENUMERA8OR]

@eyurtsev Would like to contribute to this issue. Please tell me what to do.

[jooray]

What branch are you running? I get NotImplementedError, which seems correct, because:

https://github.com/langchain-ai/langchain/blob/4b641f87ae7abcb04e2821a38d4e057b2c3fe162/libs/community/langchain_community/llms/ollama.py#L398

does not implement with_structured_output, so the parent class raises this error.