langchain-ollama (partners): allow passing ChatMessage (with custom 'role') to ChatOllama

[rylativity]

**Description: currently, ChatOllama will raise a value error if a ChatMessage is passed to it, as described here. This PR enables passing a langchain ChatMessage with a custom 'role' attribute through the langchain ChatOllama class.
Issue: resolves #30122 (also related to PR #30147 and ollama-python issue ollama/ollama#8955)
Dependencies: no new dependencies

• PR title
• PR message
• Lint and test: format, lint, and test all running successfully and passing

[vercel]

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

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Mar 21, 2025 4:05am

[Clément PAPPALARDO (lemassykoi)]

from my initial stack trace in 30122, I'm afraid that I will not be able to use create_react_agent to work with granite3.2
could you confirm ? thanks for your implication

[rylativity]

Clément PAPPALARDO (@lemassykoi) you will have to explicitly use a ChatMessage instance for the "control/thinking" message like so:

from langchain_core.messages import ChatMessage

...

    messages = {
                "messages": [
                    ChatMessage(role="control", content="thinking"),
                    (
                        "user",
                        "explain catdog",
                    )
                ]
    }

Or, for example with more context:

from langchain_ollama import ChatOllama
from langchain_core.messages import ChatMessage
from langgraph.prebuilt import create_react_agent

llm = ChatOllama(
        base_url    = "http://localhost:11434",
        model       = "granite3.2",
        verbose     = True,
        disable_streaming = True,
    )
agent_supervisor = create_react_agent(
        model        = llm,
        tools        = [],
        name         = "ReAct_Agent_WIP",
        debug        = True
    )

messages = {
            "messages": [
                ChatMessage(role="control", content="thinking"),
                {
                   "role": "user",
                    "content": "explain options trading",
                }
            ]
}

for event in agent_supervisor.stream(
    input = messages,
    # config = config,
    stream_mode = "values",
    debug = True
):
    print(event)

For now, I'd recommend using a ChatMessage instance as described above. I'm looking into whether we can have it accept a dictionary message like this:

...

messages = {
            "messages": [
                {
                    "type": "chat",
                    "role": "control",
                    "content": "thinking"
                },
                {
                   "role": "user",
                    "content": "explain options trading",
                }
            ]
}

...

[rylativity]

In order to make this also work with dict type messages it would require fixing something in langchain-core - Right now there is an implicit assumption in the _convert_to_message function in langchain_core/messages/utils.py: that 'role' and 'type' values are interchangable in dict type messages (see https://github.com/langchain-ai/langchain/blob/master/libs/core/langchain_core/messages/utils.py#L325-L328), which I don't think is the desired or intended behavior. Handling this will require a deeper look at langchain-core to avoid breaking something else, but that issue already exists and is not something that's introduced by this PR.

ccurme (@ccurme) my suggestion would be to merge this PR, which allows using ChatMessage with custom 'role' attributes with ChatOllama and closes #30122. And I can open a separate discussion and/or issue regarding how to handle 'role' vs 'type' keys in dict type messages in langchain-core (described above).

[ccurme (ccurme)]

Thanks for this, I'm happy to move forward but think it's best to keep it open until the ollama sdk actually supports control roles. Then we can merge this with a test + example use in the API reference. From what I can tell it's not yet supported by ollama: ollama/ollama-python#462

Let me know if my thinking makes sense.

[rylativity]

ccurme (@ccurme) makes sense to me.

Ollama server already accepts arbitrary roles, and the ollama-python maintainers have now expressed a preference to remove role checking/validation entirely from the python sdk. I expect ollama-python will support it shortly - as soon as this PR is reviewed, approved, and merged ollama/ollama-python#462

[rylativity]

ccurme (@ccurme) the associated functionality on ollama has now been merged into ollama (ollama/ollama#9874 (comment)) and ollama-python sdk (ollama/ollama-python#462 (comment)).

I'll get this covered with a test or two and update the API reference, let me know if there's anything else you'd want before we can merge this

[rylativity]

ccurme (@ccurme) got some wires crossed, opening a new draft PR here to keep it clean - #30411