LangExtract is an open-source Python library developed by Google AI, designed to address the challenge of extracting structured, traceable information from unstructured text using Large Language Models (LLMs) like Gemini. It aims to deliver powerful, automated extraction with inherent traceability and transparency.

The core methodology of LangExtract revolves around several key innovations:

1. Declarative and Traceable Extraction: Users define custom extraction tasks using natural language prompts (prompt_description) that articulate the desired entities, relationships, or facts. This is augmented with high-quality "few-shot" examples (examples) in the form of lx.data.ExampleData objects, which include the source text and pre-defined lx.data.Extraction objects specifying the extraction_class, extraction_text, and attributes. These examples serve as a concrete guide for the LLM, demonstrating the precise structure and content of the desired output. A foundational aspect of LangExtract is that every extracted piece of information is directly linked back to its specific span within the original source text, enabling robust validation, auditing, and end-to-end traceability of the extracted data.

2. Schema Enforcement with LLMs: LangExtract leverages the capabilities of LLMs, primarily Gemini (though compatible with others), to enforce custom output schemas, typically JSON. This mechanism ensures that results are not only accurate but also immediately usable in downstream data pipelines. The library addresses common LLM weaknesses such as hallucination and schema drift by deeply grounding the LLM's outputs. This grounding is achieved by constraining the LLM to extract information *from* the provided source text and to fit it *into* the user-defined schema, guided by both the natural language instructions and the structural patterns demonstrated in the few-shot examples. This approach significantly reduces arbitrary generation and ensures structural consistency.

3. Scalability and Visualization: For processing lengthy documents that exceed typical LLM context windows, LangExtract implements a robust scalability strategy. It efficiently handles large volumes of text by intelligently chunking the input document into smaller, manageable segments. These segments can then be processed in parallel by the LLM. After individual chunk processing, the extracted results are aggregated to form a coherent, complete output for the entire document, effectively overcoming context length limitations. For auditing and error analysis, LangExtract provides built-in interactive visualization tools. It generates HTML reports (lx.visualize) that highlight each extracted entity directly within its original context in the source document, visually demonstrating the precise location from which information was retrieved.

In practice, a typical workflow involves:
a. Defining the extraction prompt as a natural language string.
b. Providing a list of examples which are lx.data.ExampleData instances, each containing an input text and a list of expected lx.data.Extraction objects (class, text span, and attributes).
c. Invoking the lx.extract() function with the input text (or documents), the prompt, examples, and the specified model_id.
d. Saving the structured, source-anchored JSON outputs (lx.io.save_annotated_documents) and generating an interactive HTML visualization.

LangExtract's methodology, by combining declarative natural language instructions, high-quality few-shot examples, LLM-powered schema enforcement, and intelligent chunking and aggregation for scalability, provides a robust and traceable solution for structured information extraction from diverse unstructured text domains.