BasicLangExtractRecognizer silently drops provider.language_model_params (timeout, num_ctx, ...)

[lsternlicht]

Summary

BasicLangExtractRecognizer loads langextract.model.provider.language_model_params from the yaml into self._language_model_params and forwards it to langextract.extract(..., language_model_params=...). But because the same call also passes config=ModelConfig(...) (built from provider.kwargs only), langextract/extraction.py takes the elif config: branch at line 240, which does not read language_model_params — only the else: branch at line 255 does.

Net effect: any value under provider.language_model_params in the yaml — including timeout and num_ctx — is silently dropped. Ollama falls back to its 120 s default, which is the source of the confusing

InferenceRuntimeError: Ollama Model timed out (timeout=120, num_threads=None)

users see even after setting timeout: 1800 (or any other value) in the yaml. The default bundled config at presidio-analyzer/presidio_analyzer/conf/langextract_config_basic.yaml is also affected — the timeout: 240 and num_ctx: 8192 it ships with have no effect at runtime.

Reproduction

# repro_config.yaml
lm_recognizer:
  supported_entities: [PERSON]
langextract:
  prompt_file: "langextract_prompts/default_pii_phi_prompt.j2"
  examples_file: "langextract_prompts/default_pii_phi_examples.yaml"
  model:
    model_id: "gemma3:27b"
    provider:
      name: "ollama"
      kwargs:
        model_url: "http://localhost:11434"
      language_model_params:
        timeout: 1800
        num_ctx: 8192
  entity_mappings:
    person: PERSON

from presidio_analyzer.predefined_recognizers.third_party.basic_langextract_recognizer import BasicLangExtractRecognizer

r = BasicLangExtractRecognizer(config_path="repro_config.yaml")
cfg = r._get_provider_params()["config"]
print(cfg.provider_kwargs)
# {'model_url': 'http://localhost:11434'}
# ^^ timeout and num_ctx are missing — they are stored on r._language_model_params
#    but never reach the provider.

Running any slow local model (e.g. gemma3:27b or any model whose first inference takes longer than 120 s on the host) then fails with:

langextract.core.exceptions.InferenceRuntimeError: Ollama Model timed out (timeout=120, num_threads=None)

Root cause

BasicLangExtractRecognizer._get_provider_params() returns {"config": self.lx_model_config}. In langextract/extraction.py:

elif config:                                       # taken for BasicLangExtractRecognizer
    language_model = factory.create_model(
        config=config,
        examples=...,
        use_schema_constraints=use_schema_constraints,
        fence_output=fence_output,
    )
else:                                              # only here is language_model_params read
    base_lm_kwargs = {...}
    base_lm_kwargs.update(language_model_params or {})
    filtered_kwargs = {k: v for k, v in base_lm_kwargs.items() if v is not None}
    config = factory.ModelConfig(model_id=model_id, provider_kwargs=filtered_kwargs)
    language_model = factory.create_model(config=config, ...)

Only provider.kwargs survives as ModelConfig.provider_kwargs and reaches OllamaLanguageModel(**provider_kwargs), whose constructor accepts timeout= and forwards num_ctx via **kwargs to every _ollama_query() call.

Why AzureOpenAILangExtractRecognizer is not affected

Its _get_provider_params() returns {"model_id": ..., "language_model_params": ...} — no config, so lx.extract() takes the else: branch where language_model_params is actually applied. The bug is specific to BasicLangExtractRecognizer.

Why existing unit tests did not catch this

presidio-analyzer/tests/test_basic_langextract_recognizer.py:568-570 asserts:

assert "language_model_params" in call_kwargs
assert call_kwargs["language_model_params"]["timeout"] == 180
assert call_kwargs["language_model_params"]["num_ctx"] == 8192

i.e. it verifies that language_model_params is passed to lx.extract() — but never that it reaches the ModelConfig or the provider. The assertions are satisfied while the effective behaviour is broken. This issue adds regression tests that assert values arrive on ModelConfig.provider_kwargs as well.

Proposed fix

Merge self._language_model_params into self.provider_kwargs before building lx_factory.ModelConfig, using setdefault so explicit provider.kwargs: entries still win. One-line behaviour change in basic_langextract_recognizer.py. PR forthcoming.

Environment

• presidio-analyzer 2.2.362
• langextract 1.x
• Python 3.12, macOS 15.5 (Darwin 25.5.0)
• Ollama 0.x, model gemma3:27b / gemma4:31b