Previous OCR demo acting up...

[Raymond Camden]

Back in April I blogged (https://www.raymondcamden.com/2026/04/11/testing-ocr-with-chrome-built-in-ai) about testing how well the Prompt API could handle OCR. I'm doing another demo and seeing something really weird now. 

Given this CodePen: https://codepen.io/cfjedimaster/pen/dPpggYo?editors=1111

I noticed in mainline Chrome, latest, it does... nothing when the Analyze button is clicked. No errors... but it never finishes creating the session object, which in theory should be quick.

In Chrome Canary, it works, but no longer OCRs, instead it returns an array of strings that describe the image. The descriptions are fairly accurate, but it's completely missing the system prompt to look for text. 

Any ideas?

[Takashi Sato]

Hi Raymond,

Regarding the issues you're experiencing with the OCR demo, it's worth noting that since the Chrome update on February 26th, the browser's DOM began blocking local requests if certain tags or script functions are not fully specified.

Specifically, the reader.readAsDataURL(file) function is no longer handled as broadly as it was in older versions. While previous versions resolved file extensions (PDF, TXT, DOC, XML, HTML, etc.) universally, the current requirement is that these extensions must be explicitly specified within the Chrome API reader request.

This behavior has been observed in similar bugs involving Gemini's Canvas transitions for LaTeX files and PDF extensions. For further clarification on this requirement, I recommend checking the Chrome AI Building documentation about the Reader and Translate extensions.

Best regards,

Takashi Sato

--
You received this message because you are subscribed to the Google Groups "Chrome Built-in AI Early Preview Program Discussions" group.
To unsubscribe from this group and stop receiving emails from it, send an email to chrome-ai-dev-previe...@chromium.org.
To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/01dee6e6-061c-41a8-bbb1-d8328b78f036n%40chromium.org.

[Raymond Camden]

I'm sorry but I'm confused as heck. :) Are you saying that when I give the Chrome AI model the info from the image, because it doesn't know the type, it's failing some how? How would I correct this? Wouldn't I get an error instead of Chrome Canary simply ignoring my instructions and describing the image?

Also, how does this account for my main Chrome not being able to even make the session object?

--

===========================================================================
Raymond Camden

Email : raymon...@gmail.com
Blog : www.raymondcamden.com

[Thomas Steiner]

initialInputs ==> initialPrompts :-)

To view this discussion visit https://groups.google.com/a/chromium.org/d/msgid/chrome-ai-dev-preview-discuss/CAK2avF_oJ4A4fnvnKACqZV57X81rWUQT6ecdk2ViDohA-wJORg%40mail.gmail.com.

--

Thomas Steiner, PhD—Developer Relations Engineer (blog.tomayac.com, toot.cafe/@tomayac)

Google Spain, S.L.U.

Torre Picasso, Pl. Pablo Ruiz Picasso, 1, Tetuán, 28020 Madrid, Spain

CIF: B63272603

Inscrita en el Registro Mercantil de Madrid, sección 8, Hoja M­-435397 Tomo 24227 Folio 25

----- BEGIN PGP SIGNATURE -----

Version: GnuPG v2.4.8 (GNU/Linux)

iFy0uwAntT0bE3xtRa5AfeCheCkthAtTh3reSabiGbl0ck

0fjumBl3DCharaCTersAttH3b0ttom.xKcd.cOm/1181.

----- END PGP SIGNATURE -----

[Raymond Camden]

Dangit. I'm sorry. Not only that, I'm on mainline Chrome on a different machine and it works fine. The first Chrome I tested on when I started this thread has been having some really weird network issues (it will refuse to load a URL for 60 seconds and then load it instantly) so I've been meaning to delete it and reinstall - I'll make time to do so today. Thank you!