Unstructured Freemium

Document parsing API that turns messy PDFs into clean chunks for RAG

What is Unstructured

Unstructured is a document parsing platform that turns PDFs, DOCX files, scans and HTML into clean, chunked JSON that retrieval systems can actually use — layout, tables and headers preserved.

Brian Raymond and team started the project in 2022, open-sourced the Python library, and now sell a managed API plus a platform for pipelines. The company’s framing is blunt: roughly 80% of enterprise data sits in unstructured documents, and feeding raw PDF text to a retriever produces garbage.

The open-source library has 10,000-plus GitHub stars and handles the parsing itself; the paid product adds hosted throughput and connectors to S3, SharePoint and Azure Blob.

Product Features

  • Partition function — one call splits a file into typed elements: Title, NarrativeText, Table, ListItem, Header.
  • Format coverage — PDF, DOCX, PPTX, XLSX, HTML, EPUB, images, EML and plain text.
  • Open-source core — Apache-2.0 Python library, runs locally with pip install unstructured.
  • Chunking strategies — by_title chunking respects section boundaries and caps chunk size, which naive 1,000-character splits ignore.
  • Free tier — the library is free; the hosted API has a limited free page allowance before billing.
  • Pricing — hosted API is usage-based, commonly quoted near $0.003 per page for standard parsing and more for high-resolution OCR.
  • OCR support — Tesseract by default, with better accuracy from a paid vision model for low-quality scans.
  • Element metadata — each element carries page number, coordinates and parent section, so citations can point back to the source page.
  • Setup cost — local install pulls heavy dependencies (detectron2, poppler) and image builds get large; the hosted API exists for a reason.

Application Scenarios

  • Contracts and policies — split a 200-page PDF into clause-sized chunks with the section title attached.
  • Financial filings — parse tables out of 10-K reports instead of losing them to text extraction.
  • Scanned archive cleanup — OCR old document stashes before indexing them for search.
  • RAG pipelines — feed element JSON straight into LlamaIndex or LangChain loaders.
  • Compliance review — page-level coordinates let reviewers jump to the exact spot a claim came from.
  • Where it does not fit — clean, well-structured HTML or markdown sources that need no parsing at all, and image-heavy slide decks where OCR accuracy still disappoints. Budget-conscious teams also underestimate the per-page cost at hundreds of thousands of documents monthly.

Quick Summary

  • Unstructured parses PDFs, scans and office files into typed JSON elements for RAG.
  • The Python library is free under Apache 2.0; the hosted API bills per page.
  • Its by-title chunking preserves document structure where naive splitting fails.
  • Best for messy document corpora; unnecessary overhead for already-clean text sources.