LangExtract Library

1.2.1 · active · verified Thu Apr 16

LangExtract is a Python library for robustly extracting structured data from large language models (LLMs). It handles automatic chunking, multi-language support, and provides integrations with various LLM providers (e.g., OpenAI, Vertex AI, Ollama). The current version is 1.2.1, with a frequent release cadence, often introducing new providers, features, and bug fixes.

Common errors

```
unexpected keyword argument 'reasoning'
```
cause Incorrect parameter passing for `reasoning_effort` with specific OpenAI models (o1, o3, o4-mini, gpt-5) in `langextract` versions prior to 1.2.1.

fix Upgrade `langextract` to version 1.2.1 or newer (`pip install --upgrade langextract`).
```
InferenceConfigError: Could not resolve model provider for 'ollama'
```
cause Built-in model providers were not correctly loaded when specified by name in `ModelConfig` for `langextract` versions prior to 1.2.0.

fix Upgrade `langextract` to version 1.2.0 or newer (`pip install --upgrade langextract`).
```
ValueError: Required parameter: project
```
cause Missing or incorrect `project` parameter handling for Vertex AI Gemini Batch API calls in `langextract` versions prior to 1.1.1.

fix Upgrade `langextract` to version 1.1.1 or newer (`pip install --upgrade langextract`).
```
TypeError: OllamaLanguageModel.__init__() got an unexpected keyword argument 'model'
```
cause The `OllamaLanguageModel` parameter for specifying the model ID changed from `model` to `model_id` in `langextract` v1.0.4.

fix Update your `ModelConfig` for Ollama to use `model_id` instead of `model`, e.g., `ModelConfig(provider='ollama', model_id='llama2')`.

Warnings

breaking Prior to v1.0.4, `OllamaLanguageModel` used the `model` parameter. This was changed to `model_id` in v1.0.4 for consistency.
Fix: Update to `langextract>=1.0.4` and ensure your `ModelConfig` for Ollama uses `model_id='llama2'` instead of `model='llama2'`.
gotcha Using `reasoning_effort` with certain OpenAI models (e.g., o1, o3, o4-mini, gpt-5) in versions <1.2.1 would cause an `unexpected keyword argument 'reasoning'` error due to incorrect parameter passing.
Fix: Upgrade to `langextract>=1.2.1` to fix the parameter passing, or avoid using `reasoning_effort` for affected models if an upgrade is not possible.
gotcha When specifying an LLM provider by name (e.g., `ModelConfig(provider='ollama')`), versions <1.2.0 could fail to load built-in providers correctly, leading to an `InferenceConfigError`.
Fix: Upgrade to `langextract>=1.2.0` to ensure proper loading and resolution of model providers.
gotcha For users of Gemini Batch API with Vertex AI, versions <1.1.1 might have encountered a 'Required parameter: project' error due to incorrect parameter handling.
Fix: Upgrade to `langextract>=1.1.1` to ensure the `project` parameter is correctly passed to the Vertex AI Batch API.
gotcha The `debug` parameter in `extract()` defaulted to `True` in versions prior to v1.0.9, resulting in verbose output. It now defaults to `False`.
Fix: If you relied on the `debug=True` behavior, explicitly set `debug=True` in your `extract()` calls after upgrading to `langextract>=1.0.9`.

Install

pip install langextract Install stable version

Imports

extract
```
from langextract import extract
```
ModelConfig
```
from langextract import ModelConfig
```
ProviderConfig
```
from langextract import ProviderConfig
```

langextract_schema

from langextract.dataclasses import langextract_schema

LangExtractSchema

from langextract.dataclasses import LangExtractSchema

Quickstart

This quickstart defines a schema using `langextract_schema` and then uses the `extract` function to parse structured data from a given text, demonstrating a basic usage pattern. Ensure your LLM provider's API key is available (e.g., `OPENAI_API_KEY`).

import os
from langextract import extract, dataclasses

# Define a simple schema for extraction
@dataclasses.langextract_schema
class Person(dataclasses.LangExtractSchema):
    name: str
    age: int
    occupation: str

text = "John Doe is a software engineer aged 30."

# Use a dummy OpenAI key for quickstart, actual key from env
os.environ['OPENAI_API_KEY'] = os.environ.get('OPENAI_API_KEY', 'sk-dummy-key')

# Extract data using the schema
result = extract(
    text=text,
    schema=Person,
    model_name="gpt-3.5-turbo" # Or any other supported model
)

print(result.extracted_data)
# Expected: Person(name='John Doe', age=30, occupation='software engineer')

view raw JSON →