Abstract:

LangExtract is a Python library designed for extracting structured information from unstructured text documents, such as clinical notes or literary works, using Large Language Models (LLMs). Its core functionality centers on Precise Source Grounding, which maps every extracted entity back to its exact character location in the source text, and Reliable Structured Outputs, which enforces a user-defined data schema. The library is optimized for efficiency and recall in processing long documents through text chunking and parallel processing (e.g., up to 20 workers). LangExtract supports major LLMs like Google Gemini (recommended default) and OpenAI, features built-in support for local models via Ollama, and emphasizes extensibility through a community-maintained plugin provider system. The extraction process is complemented by an interactive HTML visualization tool for result review.

LangExtract Python Library Summary

• Core Purpose: LangExtract is a Python library utilizing LLMs to extract structured data (e.g., entities, relationships, attributes) from unstructured text based on user-defined instructions and few-shot examples.
• Key Feature: Source Grounding: The library ensures Precise Source Grounding, mapping every extraction to its exact location (char_interval) in the source text, essential for traceability, verification, and interactive visualization.
• Structured Output Reliability: It enforces a consistent output schema derived from provided examples, utilizing capabilities like Gemini's controlled generation to guarantee robust, structured results, often in JSON or YAML format.
• Optimized for Long Documents: The system overcomes context window limitations and the "needle-in-a-haystack" challenge using an optimized strategy of text chunking, parallel processing (up to max_workers=20), and multiple passes (extraction_passes=3) for enhanced recall.
• Flexible LLM Support:
  • Cloud Models: Supports Google Gemini (Flash and Pro versions recommended) and OpenAI (requires optional dependency).
  • Local Models: Includes built-in support for local inference via the Ollama interface, supporting models like Gemma2:2b.
• Extensibility via Plugins: Features a lightweight plugin system enabling Community Providers for other backends (e.g., AWS Bedrock, LiteLLM, Llama.cpp), which are automatically discovered via Python entry points.
• Visualization and Review: Results are saved to the .jsonl format, which can be instantly converted into a self-contained, interactive HTML file to visually inspect and review extracted entities within their original context.
• Setup and Configuration: Requires simple installation (pip install langextract) and API Key setup (via environment variables like LANGEXTRACT_API_KEY or provider-specific keys like GEMINI_API_KEY).
• Contribution Guidelines: Contributions require adherence to strict rules: small, focused Pull Requests (PRs under 150 lines preferred), mandatory formatting checks (./autoformat.sh), and the PR must link to an issue with demonstrable community support (5+ 👍 reactions required).