Docule vs Unstructured

Unstructured is an open-source library (and a hosted service) that partitions documents into elements. Docule is a managed API that turns them into ready-to-consume Markdown or JSON. This page walks through the actual trade-off: self-hosted flexibility versus operational simplicity, and where each pays off.

At a glance

Where Unstructured wins

Air-gapped and on-prem. If your documents cannot leave your VPC — bank internal, government, healthcare — Unstructured's open-source library is the honest answer. Docule is API-only; there is no on-prem option today.

Extensibility. Because Unstructured is Apache-licensed Python, you can fork it, swap the layout model, add a custom partitioner for your document type, or fine-tune on your own labelled data. Docule is a closed pipeline that you configure through parameters, not code.

Ecosystem. Unstructured has connectors for a long list of source systems (S3, SharePoint, Confluence, Salesforce…) and destinations (Pinecone, Weaviate, Elasticsearch). If ingest-and-index is your whole job, that ecosystem saves plumbing.

Where Docule wins

You don't have to run it. Self-hosting Unstructured with the "hi-res" strategy needs a GPU box, a layout model (~1 GB), an OCR engine, plus routine upgrades when new model weights ship. That is fine when it works and expensive when it doesn't. Docule is a POST endpoint — no infrastructure, no upgrade window, no oncall page at 3am because detectron2 OOM'd.

Financial-document accuracy out of the box. Docule was built on the messiest table layouts in finance: side-by-side cash flow statements, multi-currency segment reports, IFRS/GAAP reconciliations. Unstructured's general-purpose partitioner will get you close but leaves the "did the columns line up" quality bar to you.

Fixed monthly cost. €39/month for 40,000 credits is predictable. Self-hosted Unstructured is "free" until you count GPU hours, storage, and a Python dev's on-call time.

EU data residency, by default. Docule runs in Finland. GDPR conversations become one sentence: "the vendor is in the EU."

The real cost of self-hosted

The temptation of open-source parsing is that it looks free. The honest cost includes:

• A GPU worker (or several) for "hi-res" parsing — figure €150–400/month for a modest g5.xlarge-equivalent.
• Storage for models and cached PDFs.
• An engineer who owns upgrades: new unstructured-inference release, new PaddleOCR version, new torch pinning conflict.
• Quality monitoring — a self-hosted parser has no built-in metrics on "did the table extraction actually work". You'll build that.
• Oncall when the model OOMs on a 500-page PDF at 2am.

For a small team parsing thousands of pages a month, Docule at €39 is cheaper than a single day of your engineer's time. For a large team parsing millions, self-hosted starts to make sense — but only if extensibility or air-gapped operation is a real requirement.

When to pick which

Pick Unstructured if: you need on-prem or air-gapped, you have a custom document type that needs a custom parser, you already run Python inference services, or you need Unstructured's specific connectors.

Pick Docule if: you want a working parser today, you're parsing financial or EU documents, you don't want to run GPU boxes, or you need EU data residency without configuration.

Self-hosting is worth it when parsing quality is a differentiator of your product. It is not worth it when parsing is just a step on the way to embedding and search.

Try Docule on your own PDFs

The free tier is 6,000 credits per month — no credit card, no time limit. If you'd rather test without an account, use pdftotext.cc — same engine, one file at a time.

See also

• Docule vs LlamaParse
• Guide: PDF to JSON with Docule
• Guide: Extract tables from PDF
• Full API documentation