Models and config keys¶

Models¶

The chat models are built in the code and handed to the factory by name:

import os

from langchain_anthropic import ChatAnthropic
from langchain_openai import ChatOpenAI

factory = LangfuseChainFactory(
    langfuse,
    models={
        "openai": ChatOpenAI(model="gpt-5.5", api_key=os.environ["OPENAI_API_KEY"], timeout=30),
        "anthropic": ChatAnthropic(model="claude-sonnet-5-5", api_key=os.environ["ANTHROPIC_API_KEY"]),
    },
    allowed_config_keys={"model", "temperature", "max_tokens"},
)

Any LangChain chat model works. Everything that is not JSON, such as credentials, endpoints, HTTP clients and rate limiters, belongs in these constructors.

The prompt config selects a model with provider, a key of models:

{"provider": "anthropic", "model": "claude-opus-5-5", "temperature": 0.2, "max_tokens": 800}

With one model provider may be omitted. With several it is required, and a name outside models fails the build with ChainBuildError.

Allowed config keys¶

Every config key other than provider and template_format is an override of a field of the selected model. allowed_config_keys decides which overrides apply:

Value	What the config can change
`None` (default)	Nothing. The config only selects the model.
A set of keys, `{"model", "temperature", "max_tokens"}`	Those fields.
`"any"`	Every field, `base_url` and `api_key` included.

With the factory above, the config shown runs the chain on the Anthropic model with the model name, the temperature and the token limit from the config; the credentials and the timeout stay as built in the code. A key outside the set is ignored by the model but stays readable by hooks. template_format in the set lets the config pick the renderer (see Templates).

How an override is applied¶

An override never changes the model passed to the factory. It creates a new object of the same class from the fields of the selected model plus the config values, through the constructor, the way init_chat_model(configurable_fields=...) does. So:

The constructor validates the values. "temperature": "hot" fails the build with ChainBuildError, and the previous chain keeps serving.
A config key may be a field name or its alias. model sets model_name on ChatOpenAI.
Without overrides the chain runs on the model object itself.
An HTTP client the model built at construction is carried over. ChatOpenAI keeps its client, so an override of base_url or timeout changes the field but not the requests; ChatAnthropic rebuilds its client from the new fields.

Field names by provider¶

The keys are the constructor arguments of the model class, so they differ by provider. A key the class does not know is ignored, or, as with ChatOpenAI, moved into model_kwargs with a warning.

	`ChatOpenAI`	`ChatAnthropic`	`ChatGoogleGenerativeAI`	`ChatOllama`
Model name	`model`	`model`	`model`	`model`
Temperature	`temperature`	`temperature`	`temperature`	`temperature`
Token limit	`max_tokens`	`max_tokens`	`max_output_tokens`	`num_predict`