[Bug]: OpenAIResponse constructor silently overwrites user-provided temperature for O1 models

[Nazmus-Saif]

Bug Description

Summary

OpenAIResponses.__init__() unconditionally forces 'temperature=1.0for any model inO1_MODELS`, siltently discarding whatever value the caller passed in. There isno warning, no log, no raised exception, the value is just lost.

Affected versions

• llama-index-llms-openai (responses.py)

Relevant code

llama-index/llms/openai/response.py, inside __init__():

# TODO: Temp forced to 1.0 for o1
if model in O1_MODELS:
        temperature = 1.0

This runs before super().__init__(), so the caller's value is overwritten with no indication.

Why this is wrong

The caller has explicitly expressed in tent by passing a temperature value. Silently overriding it violates the priciple of least surprisee. The correct behavior is one of:

1. Raise a ValueError with a clear message explaining the constraint
2. Emit a logging.warning() so the caller at least knows what happened
3. Only force temperature=1.0 when the caller did not explicitly set it

Reproduction

from llama_index.llms.openai import OpenAIResponses

llm = OpenAIResponses(model="o1-mini", temperature=0.5, api_key="sk-...")
print(llm.temperature)  # prints 1.0 - caller's 0.5 was silently discarded

Additional context

The #TODO comment in the source is the only ack of this bahavior - it is not documented anywhere in the official docs or the class docstring.

A minimal no-breaking fix you may consider:

if model in O1_MODELS and temperature == DEFAULT_TEMPERATURE:
    temperature = 1.0  # only force if caller didn't explicitly set it

Or the stricter correct fix: raise/warn instead of silently overwriting.

Version

0.14.23

Steps to Reproduce

1. Install the package:
   pip install llama-index-llms-openai

2. Instantiate OpenAIResponses with an O1 model and a custom temperature:

   from llama_index.llms.openai import OpenAIResponses

   llm = OpenAIResponses(
   model="o1-mini",
   temperature=0.5,
   api_key="sk-..."
   )

   print(llm.temperature)

   Expected: 0.5

   Actual: 1.0 - caller's value was silently discarded

3. Check responses.py init() to confirm the override:

   TODO: Temp forced to 1.0 for o1

   if model in O1_MODELS:
   temperature = 1.0

Relevant Logs/Tracebacks

[dosubot]

Hi @Nazmus-Saif! I'm Dosu and I'm helping the LlamaIndex team.

You've correctly identified the issue. The same unconditional override exists in both base.py and responses.py [1] [2]:

# TODO: Temp forced to 1.0 for o1
if model in O1_MODELS:
    temperature = 1.0

This affects OpenAI (the main chat class) and OpenAIResponses equally — the caller's value is silently discarded with no warning or log.

A few options to fix this, in order of preference:

1. Only override when the caller didn't explicitly set temperature — your suggested approach using DEFAULT_TEMPERATURE as the sentinel is clean and non-breaking:

   if model in O1_MODELS and temperature == DEFAULT_TEMPERATURE:
       temperature = 1.0

2. Emit a warning when the override fires, so callers aren't left guessing:

   if model in O1_MODELS:
       if temperature != 1.0:
           import warnings
           warnings.warn(
               f"Model '{model}' only supports temperature=1.0; ignoring provided value {temperature}.",
               UserWarning,
               stacklevel=2,
           )
       temperature = 1.0

3. Raise a ValueError for the strictest approach — makes the constraint explicit but is a breaking change for existing code that passes a temperature.

The TODO comment suggests this was always intended as a temporary workaround. A PR fixing both base.py and responses.py at the same time would be welcome here.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.