LangExtract

LangExtract is a Google open-source information extraction library for teams that need structured data from messy text without losing the link back to the original source. It is built for developers, data teams, healthcare researchers, legal teams, and AI builders who want LLM extraction results they can review instead of blindly trusting a model response. The library takes user-defined instructions and few-shot examples, sends the task to an LLM backend, and returns structured extractions with source grounding. The core difference is traceability. LangExtract maps extracted entities back to exact character spans in the input text, so reviewers can see where every value came from. That matters for clinical notes, contracts, reports, customer feedback, research archives, and any workflow where an extracted field needs an audit trail. If a value cannot be grounded in the source, developers can detect and filter that result instead of silently mixing it into a dataset. LangExtract also handles long documents better than a single prompt pasted into a chat interface. It supports chunking, parallel processing, multiple extraction passes, and smaller context buffers for higher recall on large texts. The project includes examples for literature and medical extraction, including workflows that process full documents and then save annotated JSONL output. Results can be reviewed through a self-contained HTML visualization, which makes it easier to inspect thousands of extracted entities in context and debug prompt or schema changes. For model access, LangExtract is flexible. Public documentation describes support for Google Gemini and Vertex AI, OpenAI, Ollama and local models, plus custom provider plugins. That lets a team choose between hosted models for quality and speed, local models for privacy, or custom providers for internal infrastructure. The library is distributed on PyPI and the GitHub project is Apache-2.0 licensed, so the software itself is free to use while hosted model calls are billed by the provider a developer chooses. LangExtract is best for builders who are moving from demos to production extraction pipelines. It gives them a repeatable pattern: define the extraction task, provide examples, run against unstructured text, inspect grounded results, and iterate. It also gives reviewers and subject-matter experts a practical artifact to inspect: highlighted spans, structured records, and source text shown together instead of a detached model answer. That makes it useful for quality assurance, prompt tuning, evaluation datasets, and human-in-the-loop workflows where mistakes need to be found quickly. It is not a no-code document automation app or a managed database product. It is a Python developer library for making LLM-powered extraction more reliable, reviewable, and portable across model providers.