types: enable passing messages with arbitrary role

[rylativity]

Ollama's python client currently only permits messages with roles of either 'user', 'assistant', 'system', or 'tool'.

As described here - ollama/ollama#8955 (comment) - this prevents sending messages with the "control" role, which was introduced by IBM's granite3.2 to enable turning the "thinking" functionality on and off as needed.

This pull request adds "control" to the list of valid values for a message's "role" field/attribute so that these messages can be passed through to ollama.

Note: ollama server already handles messages with this role appropriately, but the python client currently does not allow sending them to ollama server. This resolves the issue in the python client.

[lemassykoi]

thanks, but I did your modification on my computer, and enabled full traceback.
The problem is from langchain_core/messages/utils.py at func _create_message_from_message_type

[rylativity]

This PR addresses the original issue you reported within the ollama-python client. I was able to reproduce your original issue with the original code that you shared, and it does not use langchain.

The new traceback you shared above is different and is a separate error coming from langchain-core, which is a different project/codebase.

Will you please share the code or command that you ran that resulted in the langchain error above? I'm happy to take a look and try to open a PR for langchain-core as well.

[gabe-l-hart]

One quick addition if we're going to go the route of expanding the set of literals, we should consider including "document" as well.

[lemassykoi]

This PR addresses the original issue you reported within the ollama-python client. I was able to reproduce your original issue with the original code that you shared, and it does not use langchain.

The new traceback you shared above is different and is a separate error coming from langchain-core, which is a different project/codebase.

Will you please share the code or command that you ran that resulted in the langchain error above? I'm happy to take a look and try to open a PR for langchain-core as well.

Thanks, this is my orighinal post with original code :

langchain-ai/langchain#30122 (comment)

[ParthSareen]

Hey @rylativity thanks for this PR. I think we're gonna remove the validation instead of hardcoding all the possible values. This is expected to grow with various training and post-training. Will close on merge of that

[rylativity]

@ParthSareen I'd prefer that approach as well.

I'm happy to open a PR (or modify this one) to remove the validation entirely unless there is already a PR opened for it.

[ParthSareen]

@ParthSareen I'd prefer that approach as well.

I'm happy to open a PR (or modify this one) to remove the validation entirely unless there is already a PR opened for it.

I don't think there is yet. You're welcome to take a crack at it. Maybe we can start with the Ollama repo and continue from there. Thanks!

[rylativity]

@ParthSareen I've updated this PR to allow passing arbitrary roles through ollama-python rather than checking for a value in a specific set of literals.

Ollama server already appears to allow arbitrary roles to be passed and processed. The only area where it is validated and prevented is in the Modelfile parsing. I've opened a PR to address the Modelfile parsing here - ollama/ollama#9874

Here's an example curl request to an ollama server instance without any modifications showing that it doesn't validate or block roles:

user@host ollama % curl http://localhost:11434/api/chat -d '{
  "model": "llama3.2",
  "messages":[
        {"role":"some random role","content":"who cares"},
        {"role":"user", "content":"Why is the sky blue?"}
  ], 
  "stream":false
}'

{
    "model": "llama3.2",
    "created_at": "2025-03-18T23:54:39.716608Z",
    "message": {
        "role": "assistant",
        "content": "The sky appears blue because of a phenomenon called scattering, which ..."
    },
    "done_reason": "stop",
    "done": true,
    "total_duration": 3946227417,
    "load_duration": 711783792,
    "prompt_eval_count": 31,
    "prompt_eval_duration": 180015208,
    "eval_count": 323,
    "eval_duration": 3051942458
}

[rylativity]

@ParthSareen formatting and linting issues reported by ruff should now be resolved for whenever you want to rerun the testing/linting workflow

[ParthSareen]

Will wrap this up

[ParthSareen]

Thanks for updating the tests with the change as well :)