langextract（langextract 搜索）

langextract — LLM-Powered Structured Information Extraction

Extract structured data from unstructured text with character-level provenance. Every extracted entity traces back to exact character offsets in the source document.

When to use this skill

• Extracting entities, relationships, or facts from unstructured text
• Processing clinical notes, legal documents, research papers, or reports
• Building NLP pipelines that need citation-level traceability (not just extracted values)
• Long-document extraction (chunking + parallel workers + multi-pass for recall)
• Replacing fragile regex/rule-based extraction with LLM-driven schema enforcement
• Generating interactive HTML visualizations of annotated text

1. Installation

# Standard install (Gemini backend — default)
pip install langextract

# With OpenAI support
pip install langextract[openai]

# Development
pip install -e ".[dev]"

API key setup:

export LANGEXTRACT_API_KEY="your-gemini-or-openai-key"
# Gemini keys: https://aistudio.google.com/app/apikey
# OpenAI keys:  https://platform.openai.com/api-keys

2. Core concepts

3. Basic extraction

import langextract as lx
import textwrap

prompt = textwrap.dedent("""\
    Extract characters, emotions, and relationships in order of appearance.
    Use exact text for extractions. Do not paraphrase or overlap entities.
    Provide meaningful attributes for each entity to add context.""")

examples = [
    lx.data.ExampleData(
        text="ROMEO. But soft! What light through yonder window breaks?",
        extractions=[
            lx.data.Extraction(
                extraction_class="character",
                extraction_text="ROMEO",
                attributes={"emotional_state": "wonder"}
            ),
        ]
    )
]

result = lx.extract(
    text_or_documents="Lady Juliet gazed longingly at the stars...",
    prompt_description=prompt,
    examples=examples,
    model_id="gemini-2.5-flash",
)

# Access results
for extraction in result.extractions:
    print(f"[{extraction.extraction_class}] '{extraction.extraction_text}' "
          f"@ chars {extraction.start}–{extraction.end}")

4. Long-document extraction (URL input, multi-pass, parallel)

result = lx.extract(
    text_or_documents="https://www.gutenberg.org/files/1513/1513-0.txt",
    prompt_description=prompt,
    examples=examples,
    model_id="gemini-2.5-flash",
    extraction_passes=3,   # 3 independent runs, results merged
    max_workers=20,        # parallel chunk processing
    max_char_buffer=1000   # smaller focused context windows
)
# Romeo & Juliet (147k chars / ~44k tokens) → 4,088 entities extracted

5. OpenAI backend

import os, langextract as lx

result = lx.extract(
    text_or_documents=input_text,
    prompt_description=prompt,
    examples=examples,
    model_id="gpt-4o",
    api_key=os.environ.get("OPENAI_API_KEY"),
    fence_output=True,
    use_schema_constraints=False
)

6. Local LLMs via Ollama

result = lx.extract(
    text_or_documents=input_text,
    prompt_description=prompt,
    examples=examples,
    model_id="gemma2:2b",
    model_url="http://localhost:11434",
    fence_output=False,
    use_schema_constraints=False
)

7. Visualize results

lx.io.save_annotated_documents([result], output_name="results.jsonl", output_dir=".")
html_content = lx.visualize("results.jsonl")
with open("visualization.html", "w") as f:
    f.write(html_content.data if hasattr(html_content, "data") else html_content)
# Open visualization.html in browser → color-coded annotations over source text

8. Key parameters reference

9. Custom provider plugin

import langextract as lx

@lx.providers.registry.register(r'^mymodel', r'^custom')
class MyProviderLanguageModel(lx.inference.BaseLanguageModel):
    def __init__(self, model_id: str, api_key: str = None, **kwargs):
        self.client = MyProviderClient(api_key=api_key)

    def infer(self, batch_prompts, **kwargs):
        for prompt in batch_prompts:
            result = self.client.generate(prompt, **kwargs)
            yield [lx.inference.ScoredOutput(score=1.0, output=result)]

Package as a PyPI plugin with entry point:

[project.entry-points."langextract.providers"]
myprovider = "langextract_myprovider:MyProviderLanguageModel"

Disable all plugins: LANGEXTRACT_DISABLE_PLUGINS=1

10. Use cases

Best practices

1. Write precise prompts — specify "use exact text, do not paraphrase" to keep offsets accurate
2. Use few-shot examples — 2–3 examples covering edge cases dramatically improves accuracy
3. Tune max_char_buffer — smaller values (500–1000) give more focused context; larger values reduce API calls
4. Use extraction_passes=3 for long docs — independent runs catch entities missed in single pass
5. Set max_workers — parallelization dramatically speeds up long-document processing
6. Verify offsets — result.text[extraction.start:extraction.end] must equal extraction_text
7. Use visualization — HTML output makes it easy to spot extraction errors and coverage gaps

References

• GitHub: google/langextract
• PyPI: langextract
• Google Developers Blog announcement
• Provider Plugin README
• Long-document Example
• License: Apache 2.0