extract(temperature=...) is silently ignored when config= or model= is passed #529

Summary

lx.extract() accepts a top-level temperature argument documented as "Set to
0.0 for deterministic output". That value is only ever merged into the provider
kwargs inside the model_id branch of extraction.py. When the caller supplies
a pre-built config= (a factory.ModelConfig) or model= (a language model
instance), temperature is dropped: no merge, no error, no warning.

For the OpenAI provider the consequence is visible on the wire. The provider is
constructed with temperature=None, and openai.py only sets the parameter
when it is not None:

temp = normalized_config.get('temperature', self.temperature)
if temp is not None:
  api_params['temperature'] = temp

so the request omits temperature entirely and generation runs at whatever the
server default is.

Versions affected

• Reproduced at HEAD 9853e4466c2a9c52c10280e0c6474d1c3b36dea1 (version 1.6.0).
• Also present in the released 1.x line installed from PyPI.
• Location: langextract/extraction.py, the if model: / elif config: /
  else: model-resolution block. Only the final else: branch builds
  base_lm_kwargs containing "temperature": temperature.

Repro

Self-contained, no API key and no network. A stub replaces openai.OpenAI and
records the kwargs the provider passes to chat.completions.create, which is
the wire payload.

import json, types
from unittest import mock

import langextract as lx
from langextract import factory
from langextract.core import data

CAPTURED = []

class _FakeCompletions:
  def create(self, **api_params):
    CAPTURED.append(api_params)
    payload = json.dumps({"extractions": [{"entity": "Alice"}]})
    message = types.SimpleNamespace(content=payload)
    return types.SimpleNamespace(choices=[types.SimpleNamespace(message=message)])

class _FakeOpenAI:
  def __init__(self, *a, **kw):
    self.chat = types.SimpleNamespace(completions=_FakeCompletions())

EXAMPLES = [
    data.ExampleData(
        text="Alice went to Berlin.",
        extractions=[data.Extraction(extraction_class="entity",
                                     extraction_text="Alice")],
    )
]

def run(label, **kw):
  CAPTURED.clear()
  with mock.patch("openai.OpenAI", _FakeOpenAI):
    lx.extract(text_or_documents="Bob went to Paris.",
               prompt_description="Extract entities.",
               examples=EXAMPLES, use_schema_constraints=False, **kw)
  print(label, "->", sorted(CAPTURED[0]))

run("A. model_id= + temperature=0.0",
    model_id="gpt-4o-mini", api_key="not-a-real-key", temperature=0.0)

run("B. config=   + temperature=0.0",
    config=factory.ModelConfig(model_id="gpt-4o-mini",
                               provider="OpenAILanguageModel",
                               provider_kwargs={"api_key": "not-a-real-key"}),
    temperature=0.0)

Actual output (HEAD 9853e44)

A. model_id= + temperature=0.0 -> ['messages', 'model', 'n', 'response_format', 'temperature']
B. config=   + temperature=0.0 -> ['messages', 'model', 'n', 'response_format']

Case B has no temperature key at all.

Expected

Both calls send temperature=0.0. At minimum, case B should tell the caller
that the argument is being discarded rather than accepting it silently.

The workaround that does work today is putting the value inside the config:

factory.ModelConfig(model_id="gpt-4o-mini",
                    provider_kwargs={"api_key": ..., "temperature": 0.0})

but nothing in the extract() signature or docstring suggests the top-level
argument stops working once config= is supplied.

Impact

This fails silently and in the direction that is hardest to notice: the run
still succeeds and still returns plausible extractions, they are just sampled at
the server default temperature instead of the requested one.

We hit this chasing a reproducibility discrepancy against a batch-invariant vLLM
deployment. Raw API probes at temperature 0 returned byte-identical completions
across runs, but the same prompts driven through langextract with a config=
produced different extractions run to run. The missing wire parameter was the
whole cause. Because temperature is the standard lever for determinism, anyone
using config= or model= for evaluation harnesses, response caching, or
reproducibility claims is affected without any signal that their setting never
took effect.

config= is also the documented route for custom and OpenAI-compatible
providers, so the paths most likely to be used for self-hosted deterministic
setups are exactly the ones where the parameter is dropped.

Suggested fix

When config= is given, merge an explicit temperature into
config.provider_kwargs, with the explicit argument taking precedence over any
value already in the config, and without mutating the caller's ModelConfig.

When model= is given the instance is already constructed, so there is no
provider-agnostic way to apply the value; a UserWarning there matches the
existing treatment of use_schema_constraints in the same branch.

Happy to send a PR; a patch with tests is ready.