leonanpereirapinto/ai-utils/aeonbridge-langextract

LangExtract Skill Snapshot

This is a local consultation copy adapted from the upstream skill and the official project README. Keep the upstream links above as the source of truth for future updates.

When To Use

Use LangExtract when you need to:

• extract structured entities from unstructured text
• keep every extraction tied to the exact source span
• process long documents that exceed normal prompt sizes
• build few-shot extraction pipelines without fine-tuning
• generate JSONL outputs and HTML visualizations for human review

Strengths

• source grounding for traceability and validation
• structured outputs driven by prompt instructions and examples
• long-document handling with chunking, parallel workers, and extraction passes
• support for Gemini, OpenAI, Vertex AI, and Ollama-style local setups
• practical review flow with saved annotations and HTML visualization

Recommended Workflow

1. Write a precise prompt_description describing what to extract.
2. Add a few high-quality ExampleData examples with grounded Extraction items.
3. Run lx.extract(...) on the target text or documents.
4. Review the extracted spans and attributes.
5. Save results to JSONL and visualize them in HTML when manual review matters.

Quick Start

import langextract as lx
import textwrap
prompt = textwrap.dedent("""\
    Extract all medications mentioned in the text.
    Include name, dosage, and frequency.
    Use exact text from the source.""")
examples = [
    lx.data.ExampleData(
        text="Patient prescribed Lisinopril 10mg daily.",
        extractions=[
            lx.data.Extraction(
                extraction_class="medication",
                extraction_text="Lisinopril 10mg daily",
                attributes={
                    "name": "Lisinopril",
                    "dosage": "10mg",
                    "frequency": "daily",
                },
            )
        ],
    )
]
result = lx.extract(
    text_or_documents="Patient continues Metformin 500mg twice daily.",
    prompt_description=prompt,
    examples=examples,
    model_id="gemini-2.5-flash",
)
for extraction in result.extractions:
    print(extraction.extraction_text, extraction.attributes)

Common Patterns

Save JSONL And Visualize

import langextract as lx
lx.io.save_annotated_documents(
    [result],
    output_name="extractions.jsonl",
    output_dir=".",
)
html = lx.visualize("extractions.jsonl")
with open("extractions.html", "w") as f:
    if hasattr(html, "data"):
        f.write(html.data)
    else:
        f.write(html)

Long Documents

result = lx.extract(
    text_or_documents=long_text,
    prompt_description=prompt,
    examples=examples,
    model_id="gemini-2.5-flash",
    extraction_passes=3,
    max_workers=20,
    max_char_buffer=1000,
)

OpenAI Models

result = lx.extract(
    text_or_documents=text,
    prompt_description=prompt,
    examples=examples,
    model_id="gpt-4o",
    use_schema_constraints=False,
    fence_output=True,
)

Local Models Through Ollama

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

Practical Guidance

• Keep examples small, realistic, and consistent with the output you want.
• Prefer grounded extraction_text values copied directly from the source text.
• Start with one simple example before scaling to many classes or attributes.
• Increase extraction_passes when recall matters more than speed.
• Use visualization whenever the extraction pipeline needs auditing.
• Filter out ungrounded results if your workflow requires strict traceability.

Key Parameters

• model_id: target model, such as gemini-2.5-flash or gpt-4o
• examples: few-shot examples that define the extraction style and schema
• extraction_passes: extra passes to improve recall on long text
• max_workers: parallel worker count for larger workloads
• max_char_buffer: overlap and context sizing for long-document chunking
• use_schema_constraints: disable when using providers that do not support the same constraints
• fence_output: often useful with OpenAI models

Helpful Links

• Official repository: https://github.com/google/langextract
• Examples: https://github.com/google/langextract/tree/main/examples
• Docs examples: https://github.com/google/langextract/tree/main/docs/examples
• Gemini API docs: https://ai.google.dev/