Langextract — Structured Extraction from Text Using LLMs by Google

Introduction

Langextract is a Python library developed by Google for extracting structured data from unstructured text using large language models. It provides a schema-driven API that returns typed objects with source grounding, so every extracted field links back to the specific text span that supports it.

What Langextract Does

• Extracts structured records from free-form text using LLM reasoning
• Supports user-defined schemas using Python dataclasses or dictionaries
• Provides source grounding linking each extracted field to its origin span
• Offers interactive visualization of extraction results and provenance
• Works with Gemini models and supports custom LLM backends

Architecture Overview

Langextract wraps LLM calls with a schema-aware prompt construction layer. User-defined schemas are compiled into extraction instructions that guide the model to produce structured JSON output. A post-processing layer validates types, resolves references, and computes character-level grounding spans. The library uses Gemini API by default but exposes hooks for plugging in alternative model providers.

Self-Hosting & Configuration

• Install via pip with Python 3.10 or later
• Set the GOOGLE_API_KEY environment variable for Gemini access
• Define extraction schemas as Python dataclasses or typed dicts
• Configure model parameters like temperature and max tokens
• Optional visualization server runs locally for debugging extractions

Key Features

• Schema-driven extraction with automatic type validation
• Source grounding traces every field back to its supporting text
• Batch processing for extracting from multiple documents efficiently
• Interactive visualization dashboard for reviewing extraction quality
• Supports nested schemas and list-valued fields

Comparison with Similar Tools

• Instructor — structured LLM outputs via Pydantic; Langextract adds source grounding
• Outlines — grammar-constrained generation; Langextract operates at a higher schema level
• BAML — type-safe AI functions; Langextract focuses specifically on extraction with provenance
• Docling — document parsing; Langextract handles unstructured text to structured data

FAQ

Q: Which models does it support? A: Gemini models are supported by default. Custom backends can be configured via the provider interface.

Q: What is source grounding? A: Each extracted field includes character offsets pointing to the exact text span that the model used as evidence.

Q: Can it process PDFs or HTML? A: Langextract operates on plain text. Pair it with a document parser like Docling for PDF input.

Q: Is it suitable for production workloads? A: Yes, it includes batch processing, retry logic, and structured error handling for production use.

Sources

• https://github.com/google/langextract