Unstructured

What it does

Unstructured is the boring step before retrieval: it splits local files into elements — headings, tables, paragraphs — and strips the layout noise that makes a parsed PDF useless to a model. It handles scans and mixed-format folders rather than URLs.

It is a library and an API, not a chat interface. You still need somewhere to put the output and a reason to index it.

Standout features

• PDF, scan, HTML and image partitioning
• Table and layout structure preserved
• Connectors for common vector stores
• Self-hostable open-source core

Main job: turn messy local documents into clean structured elements before they reach a model or a vector store.

Best for

Use it when supplier specs, scanned certificates or old quotation PDFs need to become searchable text for an internal index.

Getting started

Point it at one folder first, inspect the extracted elements on twenty files, then tune the table and OCR settings before running the whole archive. Cache the output so re-indexing does not re-parse everything.

Watch out for

Watch out: customer drawings and certificates are usually covered by an NDA, and OCR sends them somewhere. Check where processing happens, redact logos on samples, and verify table extraction on one real spec sheet before trusting any numbers.