MILabs
Tools

olmOCR PDF text extraction for self-hosted document work

olmOCR is open-source software that converts PDFs and scanned pages into text, run on infrastructure the business controls.

Updated 2026-05-16

What olmOCR is

olmOCR is an open-source toolkit from the Allen Institute for AI, known as AI2, for turning PDFs and scanned document images into clean plain text. It uses a vision language model to read each page, so it can handle layouts that simpler text extractors struggle with, such as multi-column pages, tables and mixed text. The project is published openly on GitHub, so a business can inspect the code and run it on hardware it controls. Large collections of documents are where it tends to earn its keep, and the work is typically run on GPU hardware. For a business weighing PDF text extraction, the tool produces text and leaves the decisions to other software or a person. Businesses with long archives of PDFs, scanned forms or reports often use it as the first step in a larger pipeline that searches, summarises or files the content. The text it produces is the raw material for search, filing and question and answer tools.

Allen Institute for AI: official site

olmOCR in plain language

Optical character recognition has existed for decades, yet older tools tend to read a page as a flat stream of characters and struggle with headings, columns, footnotes and stamps. olmOCR treats each page as an image and asks a model to rebuild its text and structure. A scanned lease with a sidebar, or a supplier statement with merged table cells, is therefore more likely to come out usable.

Running it yourself changes what the tool means for your data. An IT contractor can read the code, pin a version and choose where it runs, whether on a server in your office or a rented machine with a capable graphics card. Documents need not leave that environment to be converted. The cost is responsibility, since someone must keep the software running, watch disk space and check output when new scans look different.

Expect the output to be raw material rather than a finished record. Handwriting, low-resolution faxes, skewed scans and heavy stamps still reduce accuracy, so any figure that matters should be checked against its source. Text from olmOCR usually feeds other systems, such as search, spreadsheet extraction or a question and answer layer, and those systems shape what it is used for.

What businesses use it for

  • Converting long archives of scanned contracts and reports into searchable text
  • Pulling text from supplier invoices and receipts before they are filed
  • Preparing policy documents and manuals for a staff question and answer assistant
  • Digitising paper forms so their details can be entered into a CRM or accounting system
  • Keeping document text on infrastructure the business controls

Who it suits

  • Firms holding large volumes of PDFs or scanned pages
  • Teams that must keep document content on machines they own
  • Owners who want an open-source base they can inspect and maintain

How MILabs sets it up

MILabs begins with the owner or team lead, deciding which document types matter, where the source files are kept and who may read the extracted text. Because olmOCR is self-hosted, the machine it runs on is chosen first, sized to the document volume and to the hardware the business can look after. MILabs then connects olmOCR to the folders, storage and document systems the business already uses, so files come in and text goes out without manual copying. Permissions come next, so only the right people and services can reach the output. Finally, the owner and MILabs agree who maintains the installation, applies updates and checks results when document formats change.

Builds it fits

Industries it fits

olmOCR in practice

Inside a business, olmOCR usually sits at the front of a short chain rather than on a screen staff open daily. A batch of scanned files lands in an intake folder, the conversion job reads each one, and the text appears beside it in an output location with a predictable name. Someone then decides whether it feeds a search index, a contract register or a spreadsheet mapping step.

The people who notice most are usually those handling exceptions. A rotated page, a stamp across a signature block or a fax photocopied several times can come through with gaps, and those files drop into a review queue. A sensible rule picks which items get a second look, such as any file returning very little text or any figure feeding a payment.

Capacity is the other practical question. Conversion is heavy work for a graphics card, so batches are usually queued and the team agrees which folders take priority. Access follows the output rather than the input, so a text file from a personnel record should be readable only by people who could open the original scan. With several sources, each intake path stays separate so one fault cannot spread.

good questions, straight answers

Frequently asked questions

Which builds work well with olmOCR?
Builds that begin with a document and need its text suit olmOCR best. A private document question and answer server is a natural fit, because the text stays on the business's own machine. Invoice capture and document filing automation also benefit, since clean text makes fields easier to find and route. Each build needs checks, as extracted text is only as accurate as the scans behind it.
What is olmOCR?
olmOCR is an open-source toolkit from the Allen Institute for AI that converts PDFs and document images into plain text. It uses a vision language model to read each page, which helps with complex layouts. The code is published on GitHub, so a business can run it on its own hardware rather than sending documents to a hosted service. It is a text extraction tool that sits inside a larger workflow.
How does MILabs set up olmOCR?
MILabs starts with the owner or team lead to decide which documents matter and where the output should go. Because olmOCR is self-hosted, the machine it runs on is chosen first. The team then connects it to the folders or systems the business already uses, sets permissions on the extracted text and agrees who maintains the installation. The result is a setup the business understands and can operate after handover.
Where does olmOCR store data?
olmOCR is self-hosted, so the documents it reads are processed on the machine where it is installed rather than by a hosted service. The extracted text is written to the output location the business chooses, such as a local folder or a document system. MILabs sets those locations during setup and records them for the owner. Storage needs depend on the volume of files, so plan disk space with that in mind.
When is olmOCR the wrong choice?
olmOCR is the wrong choice when a business has no suitable hardware or no one to maintain a self-hosted installation. It is also unnecessary for simple, born-digital PDFs that a standard text export already handles well. If the team expects a hosted service with no infrastructure to manage, another option will suit better. In those cases, MILabs can talk through the alternatives with the owner before any work begins.
let's make something happen

Workflow automation: how we can help

Choose how you want to work on this with MILabs. Both options start with an enquiry.

Discovery is free, no obligation from the first chat.

Let us build it for you

We build it and look after it.

Build it yourself, we'll help

You build it with our direction, teaching and a weekly check-in.