Summary

When using Gemini 3 models (gemini-3-flash-preview, gemini-3-pro-preview) with LangExtract, the API calls timeout because Gemini 3 defaults to thinking_level: high which adds significant latency. There's currently no way to pass thinking_config to reduce this latency.

Problem

The _API_CONFIG_KEYS allowlist in langextract/providers/gemini.py (line 40-48) does not include thinking_config:

_API_CONFIG_KEYS: Final[set[str]] = {
    'response_mime_type',
    'response_schema',
    'safety_settings',
    'system_instruction',
    'tools',
    'stop_sequences',
    'candidate_count',
}

This means any thinking_config passed via language_model_params gets filtered out (line 186-188):

self._extra_kwargs = {
    k: v for k, v in (kwargs or {}).items() if k in _API_CONFIG_KEYS
}

Impact

• Gemini 3 Flash (designed for speed) times out on simple extraction tasks
• Users cannot set thinking_level: "minimal" to reduce latency
• Forces users to use older models like gemini-2.5-flash instead of newer Gemini 3 models

Proposed Solution

Add thinking_config to _API_CONFIG_KEYS:

_API_CONFIG_KEYS: Final[set[str]] = {
    'response_mime_type',
    'response_schema',
    'safety_settings',
    'system_instruction',
    'tools',
    'stop_sequences',
    'candidate_count',
    'thinking_config',  # Add this for Gemini 3 support
}

This would allow users to pass:

result = lx.extract(
    text_or_documents=text,
    prompt_description=prompt,
    examples=examples,
    model_id="gemini-3-flash-preview",
    language_model_params={
        "thinking_config": {"thinking_level": "minimal"}
    }
)

References

• Gemini 3 Developer Guide - thinking_level parameter
• Gemini Thinking Documentation

Environment

• LangExtract version: 1.1.1
• Python: 3.12
• Models tested: gemini-3-flash-preview, gemini-3-pro-preview