2020051 - [wpt-sync] Sync PR 58120 - Create renderer side code for ai taxonomy API

Status: NEW

(Testing :: web-platform-tests, task, P4)

Description

[Web Platform Test Sync Bot [:wpt-sync] (Matrix: #interop:mozilla.org)]

Sync web-platform-tests PR 58120 into mozilla-central (this bug is closed when the sync is complete).

PR: https://github.com/web-platform-tests/wpt/pull/58120
Details from upstream follow.

devincabillo <[e-mail]> wrote:

Create renderer side code for ai taxonomy API

Implementation stubbed out in accordance to go/chrome-ipc-stubs-trial

Feature currently disabled and returns an empty promise
pending browser side implementation and link

What is this API supposed to do?

• It will be used in devTrial for adds to perform Taxonomy on Webpages. Strings will be passed through this layer (blink) and then end up on the ai engine within optimization_guide to classify text into IAB taxonomies using a tinyGemma model.

Who will call it and who will implement it?
-It should be an open endpoint eventually for developers, but for now it will be a built in API to pass text to and provide IAB Taxonomies. Browser process will eventually pass the logic to optimization_guide to perform inference

What data will be passed where and by whom?
-For the DevTrial we will be focused on the developer(Ads) passing a simple text string (the input to be classified) and an optional dictionary of settings

If text is being parsed (it sounds like it is), where does that occur?
-In the Blink (Renderer) layer, the text is treated strictly as a string. No linguistic parsing or tokenization occurs in the rendering engine. The string is serialized and passed across the Mojo IPC boundary to the Browser process. The actual "parsing" (tokenization, embedding, and model inference) happens safely inside the underlying AI execution engine

Bug: 483707964
Change-Id: I93f53ebb5d06c86a84a21f104b2e53ff78d8dedf

Reviewed-on: https://chromium-review.googlesource.com/7568249
WPT-Export-Revision: 1d778fe07864e7e209e14ec169a2f51e2bd405ce