Does the text in a picture come back?

The short answer

Most of it comes back. Not where you expect, and not all of it.

txtfetch reads a PDF page with OCR when the page has 10 typed characters or fewer. A page with more is not read. A scan with a typed footer is lost. A picture on a typed page is lost.

txtfetch reads every picture in a Word, PowerPoint or Excel file. The text lands at the end of the document, after a line such as image1.png. The alt text is lost.

HTML and email keep the alt text. The picture text follows it.

The flag metadata.ocr is false in all of these cases. Do not use it to find out if OCR ran.

The engine is tika-server-standard 3.3.1.

Source: terraform/lambda-layers/build/versions.env (versions.env: TIKA_VERSION), accessed 2026-10-18.

What comes back, case by case

We sent each file to the production API. Every picture holds the words PICTURE WORDS ALPHA and a second line. Every scan holds the words SCANNED PAGE BRAVO. Every row shows its source. The raw responses are in the repo.

What to do about it

A picture on a typed PDF page is lost

Render that page to a PNG and send the PNG. A picture file is always read by OCR. Find the pages at risk first with the PDF text check.

A scan with a footer or stamp is lost

The typed footer pushes the page over 10 characters. Render the page to a PNG and send that. Then the footer is read as part of the picture.

Picture text in an Office file lands at the end

Split the text on the image file name line. The text before the first name is the document. The text after each name belongs to that picture. The snippet below does it.

Do not read metadata.ocr

It says false when OCR ran inside a PDF or an Office file. It says true only for a picture file you send on its own. Check the text, not the flag.

The PDF text check finds pages with no text layer. Run it before you ingest a batch.

Split the picture text from the document text

The picture text starts after a line such as image1.png. This snippet puts each picture under its file name. In your own code, text is the extracted_text field of the response.

Node.js, run on the docx response

import { readFileSync } from "node:fs";

const text = JSON.parse(readFileSync("responses.json", "utf8")).docx.extracted_text;

const parts = text.split(/^(image\d+\.png)$/m);
const document = parts[0].trim();
const pictures = {};
for (let i = 1; i < parts.length; i += 2) {
  pictures[parts[i]] = parts[i + 1].trim();
}
console.log(JSON.stringify({ document, pictures }));

It printed

{"document":"Body text CHARLIE before the picture.\n[image: ]\nBody text ECHO after the picture.","pictures":{"image1.png":"PICTURE WORDS ALPHA\nRevenue grew 4471 units"}}

The file name is a label. It does not tell you where the picture sat on the page.

What we did not test

We tested uploaded files with printed text in the pictures. We did not test handwriting in a picture. Read Handwriting for that. We did not test a page that txtfetch fetches from a URL.

The flag metadata.ocr reports false when OCR ran inside a PDF or an Office file. This is a known gap in the API.

?quality=premium is sold as a tier, but it does not run today. VLM_ENABLED defaults to false, and no operator has switched the tier on in production. A premium request downgrades to the standard Tika path instead.

Read the PDF page, the scanned PDF page, the Word page and the image page for the rest of each format.

Frequently asked questions

Does the text in a picture come back?

Most of it does. A scan, or a picture in Word, PowerPoint, Excel, HTML or email, is read by OCR. A picture on a typed PDF page is not. A PDF page with more than 10 typed characters is not read either.

When does txtfetch read a PDF page with OCR?

It decides page by page. A page with 10 typed characters or fewer is read by OCR. A page with more is not. A typed cover page and a scanned body both come back.

Why is the text of my scanned page missing?

The page probably holds a typed footer, a page number line or a stamp of more than 10 characters. Then only the typed text comes back. Render the page to a PNG and send the PNG.

Where does the text from a picture in a Word file appear?

At the end of the document, after a line with the picture file name, such as image1.png. In Markdown the name is a heading. In JSON it is a heading element.

Is the alt text of a picture kept?

In HTML and email, yes. It comes back in an [image: ] marker. In Word, PowerPoint and Excel, no. Word shows an empty marker.

Does metadata.ocr tell me that OCR ran?

No. It said false in every test except a picture file sent on its own. Do not use it to detect OCR. Read the text.