Unstract tutorial: invoice extraction and JSON schemas

This tutorial walks through invoice extraction in Unstract's Agentic Prompt Studio. The presenter uploads three invoices with different layouts, configures the project's LLM settings and generates a JSON schema. The example includes company details, billing information, invoice dates, tax and totals. Users can edit the schema or remove fields before generating an extraction prompt.

The presenter contrasts this workflow with fixed templates and conditional extraction rules, which can fail when vendors change document layouts. He also distinguishes Agentic Prompt Studio's automatic prompt generation from Prompt Studio, where users write their own prompts. The demonstration includes PDF-to-text conversion, a verifier run, processing all documents and an option to export data. The presenter describes using generated prompts through Unstract's API in another project.

For readers considering a self-hosted deployment, the speaker points to an open source repository and describes cloning it and running the platform on localhost. The practical walkthrough uses an account-based dashboard; it does not demonstrate a local LLM or establish offline operation.

A separate LLMWhisperer example shows invoice text retaining its spatial layout, including table placement. The dashboard reports 100% project accuracy for the demonstration, but that result does not establish accuracy across other documents.