What Is an OCR File? The Forms Recognized Text Takes

An OCR file is any file that holds text a computer recognized from a picture of a page. No file format goes by that name. Most often it is a searchable PDF, and sometimes a plain text, Word or data file. You have one when the text inside came from recognition and not from typing.

The phrase turns up in emails from opposing counsel, in vendor quotes and in clerks' notices, and it rarely comes with a definition. This post pins it down, shows the shapes it takes, and gives you ways to tell whether the file in front of you has already been through recognition.

What do people mean by an OCR file?

People use "OCR file" for a file that carries machine-recognized text, and they mean one of three things by it: a PDF with a text layer added, a text file made from a scan, or, loosely, a scan that is waiting to be recognized. Context usually tells you which.

There is no .ocr extension in everyday office use and no standard that defines the term. It is shorthand, and the three senses lead to different work.

The third row is the confusing one. Some people call any scanned document an "OCR file" because OCR is what they associate with scans. If a request could be read two ways, reply with a concrete question: "Do you want searchable PDFs, or separate text files alongside the images?"

A request for an OCR file of a deed usually means the picture plus words you can find: each party, a marginal notice, and the recording citation. A bare text file keeps those words and loses the page. Search the citation and see that it still sits with the facts the deed recites.

Phrases such as "make OCR PDF" and "convert PDF to OCR PDF" belong to the first sense. They describe taking a scan and adding recognized text to it. The guide to making a scanned PDF searchable covers that task from start to finish.

The four forms an OCR file takes: PDF, text, Word and data

Recognized text is delivered in four broad forms: a PDF that keeps the page picture and adds text, a plain text file, a word-processing document, and structured data such as a spreadsheet or JSON. The recognition is the same in each. What differs is how much of the page comes along with the words.

PDF

This is the form most people have in mind. The file still shows the scan, and a reader program can search it. It is one member of the family, and the post on the searchable PDF describes what is inside it and how to spot one.

Text

A plain text file is the simplest possible record of what the software read. It opens anywhere and hides nothing, which makes it the easiest form to compare against the page. It is also the form litigation tools exchange alongside page images.

Word

A word-processing export tries to rebuild the document so you can edit it. It is the right form when you need to revise or reuse language, and the wrong one to treat as the record, because a rebuilt page is an interpretation.

Data

Data forms carry the text in a structure a program can use. Some also carry position: where on the page each word was found. The command-line documentation for the open-source Tesseract engine shows the range in one place. Its guide to command-line usage says "The default output format is text", describes the PDF option as one that "creates a pdf with the image and a separate searchable text layer with the recognized text", and lists hOCR and TSV outputs as well.

One upload can yield more than one form. Schema Law, for example, exports the same parsed document as Word, searchable PDF, plain text, Markdown, HTML, Excel, CSV or JSON. The export formats section of the homepage says what each holds, and all eight are on every plan.

Signs that a file has already been through OCR

A file has been through OCR if it shows a scanned picture and yet lets you select words, or if it arrives with a text file of the same name beside it. Smaller clues are in the document properties, where the software that produced the file is often recorded.

No single sign is conclusive. Use several.

Reading the producer field

A PDF can record two names in its properties: the application that created the original document, and the producer, meaning the software that wrote the PDF. On a file saved from a word processor, those fields usually name the word processor or its PDF writer. On a copier scan, they tend to name the device or its scanning software. When a scan's producer names a recognition program, that is strong evidence the text layer came from it.

Treat the field as a lead, not as proof. It can be blank. It can be stripped on purpose. And the last program to save the file may overwrite it, so a scan that was merged, stamped or compressed after recognition may name the merging tool instead.

The opposite conclusion

If none of the signs is present, and no word on the page will select, the file is a plain scan. Whether to recognize it, and what to do when only some pages have text, is the subject of the post on OCR for PDF documents.

What does "OCR'd" mean in a production request or court notice?

"OCR'd" means the text of a scanned or image document has been recognized and is being delivered with it. In a production request that can mean a text file per document or searchable PDFs. In a court notice it can mean the filing must be text-searchable. The governing document decides which.

The word is a verb made from an abbreviation, and requests use it loosely. "All paper documents shall be scanned and OCR'd" tells you recognition is required. It does not tell you where the text goes. Read on for the delivery term in the same document.

One published model shows a common pattern. The EDRM Production Standards provide that each "non-searchable file containing text has a multipage OCR text file named with the DocID of the corresponding file." In that scheme, "OCR'd" is satisfied by a text file, not by a text layer inside an image.

Your own protocol may differ. Before you promise or accept "OCR'd documents", settle these points:

• Form. Searchable PDFs, or images with separate text files?
• Level. One text file per document, or per page?
• Scope. Every document, or only those with no text of their own?
• Redactions. Text made from the redacted image, so the hidden words are not delivered?
• Quality. Any stated expectation, or simply "as recognized"?

On scope, productions usually treat text copied out of an electronic file differently from text recognized from a picture. The post on e-discovery OCR explains extracted text versus OCR text, and why the two should not be confused.

Court notices are a separate matter. Filing rules differ from court to court, and a notice that a document "must be OCR'd" is best read against that court's own rule. The post on searchable PDFs for court filing collects examples. Nothing here is legal advice on what your order or rule requires.

Sidecar text files: OCR text stored beside the page image

A sidecar text file is a separate file that holds the recognized text of a document whose pages are stored as images. The two travel together and are matched by name. The image is what you look at, and the sidecar is what the search tool reads.

The EDRM standard gives the naming rule in one line: "DocID = ABC0000123; Filename = ABC0000123.txt." A folder built that way looks like this for two fictional documents:

IMAGES/
  PRICE0000101.pdf
  PRICE0000117.pdf
TEXT/
  PRICE0000101.txt
  PRICE0000117.txt

PRICE0000101.pdf might be a sixteen-page scanned supply agreement. PRICE0000101.txt then holds the recognized text of all sixteen pages in one file.

Why keep text outside the document?

Review platforms load text into a search index, and a plain file is the easiest thing to load. Page images in formats such as TIFF cannot hold a text layer the way a PDF can, so the text has to live somewhere else. A sidecar is also easy to replace when recognition is run again.

What goes wrong with sidecars

• The pair gets separated. Rename or move the image without the text file and the link is gone. Where a load file records the path to each text file, that record must be updated too. The post on the Concordance load file format shows where the path is kept.
• The text goes stale. If a page is redacted or replaced after recognition, the old sidecar still holds the earlier words.
• Page breaks vanish. A multipage text file may or may not mark where each page ends. Without marks, you cannot go from a hit to a page number.
• Characters are mangled. A text file saved in one encoding and opened in another turns accents and symbols into nonsense.

Some sidecars hold more than words. Formats such as hOCR and TSV also record where each word sat on the page, which lets a viewer highlight a hit on the image.

Which program do you need to open an OCR file?

You need no special program. An OCR file opens in whatever opens its format: a PDF reader for a searchable PDF, a text editor for a .txt, a word processor for a .docx, a spreadsheet for .xlsx or .csv. Reading recognized text and producing it are different jobs.

The search phrase "PDF OCR reader" blurs two things. A reader displays a PDF and searches text that is already there. Recognition creates that text. Some PDF programs do both, and a free reader typically does only the first. If your reader finds nothing in a scan, the reader is not broken; the file has no text yet.

Two follow-up questions come up often.

Can you open the hidden text of a PDF on its own? Yes, roughly. Select all the text on a page, copy it, and paste it into a blank document. What appears is the text layer, and reading it is the quickest quality check there is.

Can you view recognized text before choosing a form? That depends on the tool. In Schema Law the parsed pages are shown in the browser after upload, so you can read them against the original first and export afterwards. The pricing page lists the eight export formats, which every plan includes.