LLM extraction that points to exactly where it found everything

What it does

LangExtract is a Python library that feeds unstructured text—clinical notes, novels, reports—to an LLM and returns structured extractions like characters, medications, or relationships. You define the task with a short prompt and a few examples, and the library enforces a strict output schema. For long documents, it chunks the text, processes chunks in parallel, and can run multiple passes to improve recall.

The interesting bit

The library treats source grounding as a first-class concern: every extraction is mapped to exact character intervals in the original text, and if the model hallucinates something from your few-shot examples instead of the input, the result gets a char_interval = None flag so you can filter it out. It also spits out a self-contained interactive HTML file that highlights thousands of entities in context—no frontend build required.

Key highlights

• Precise source grounding: every extraction maps to exact character intervals in the original text
• Long-document handling: chunks text and runs parallel extraction passes for higher recall on books or large reports
• Flexible backends: works with Gemini, OpenAI, or local models via Ollama, plus Vertex AI Batch API for cost savings at scale
• Controlled generation: enforces structured output schemas on supported models rather than hoping JSON parses correctly
• Domain-agnostic: define any extraction task with a few examples and a prompt—no fine-tuning necessary

Caveats

• Cloud-hosted models require API keys, and the README suggests a paid Gemini tier for production throughput to avoid rate limits
• Gemini models have defined retirement dates, so you’ll need to track model lifecycle documentation to avoid surprises

Verdict

Worth a look if you need auditable entity extraction from long documents without training a custom model. Skip it if your current regex pipeline is already bulletproof—this is still an LLM doing probabilistic guesswork under the hood.