---

name: extraction description: "Build robust LangExtract extraction pipelines with examples, schemas, prompt validation, resolver tuning, chunking, and tokenizers." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

LangExtract Extraction

Use this sub-skill when the task is to design, write, review, or debug calls to lx.extract() and the core data objects that make extraction reliable and source-grounded.

Load first

• Read references/workflows.md for extraction input shapes, prompt/example contracts, output schemas, validation, resolver parameters, tokenizers, and result handling.
• Read references/troubleshooting.md when an extraction call fails, returns ungrounded spans, has schema/fence conflicts, or behaves poorly on long or non-spaced-language documents.
• Use scripts/basic_extraction.py as a safe starter script for example-based extraction.
• Use scripts/output_schema_extraction.py as a safe starter script for schema-constrained extraction.

Best-fit tasks

• Create basic or relationship extraction prompts with lx.data.ExampleData and lx.data.Extraction.
• Choose between plain text input, trusted URL string input with fetch_urls, and iterable lx.data.Document inputs.
• Tune long-document parameters such as max_char_buffer, batch_length, max_workers, extraction_passes, and context_window_chars.
• Use lx.schema.extraction_item_schema() and lx.schema.extractions_schema() with output_schema.
• Enforce prompt/example quality with PromptValidationLevel and prompt_validation_strict.
• Tune resolver alignment and parsing through resolver_params.
• Select RegexTokenizer versus UnicodeTokenizer and filter grounded AnnotatedDocument results by char_interval and AlignmentStatus.

Route elsewhere

• Provider credentials, model routing, Gemini/OpenAI/Ollama/Vertex settings, provider kwargs, and batch APIs belong in providers.
• JSONL save/load and HTML visualization belong in visualization.
• Custom provider package authoring belongs in provider-plugins.

Operating checklist

1. Define the extraction classes and attributes before writing examples.
2. Make every example extraction_text a verbatim substring of the example text and keep extractions in source order.
3. Add examples unless an explicit output_schema is present; examples are still useful with schemas because they teach the prompt.
4. Validate examples in development with PromptValidationLevel.ERROR and prompt_validation_strict=True.
5. For long documents, start with small samples and scale max_char_buffer, batch_length, max_workers, and extraction_passes deliberately.
6. Treat char_interval is None as ungrounded output until diagnosed or intentionally filtered out.