Skip to content

Tracking: Implement Embedding Support #962

Open
4 of 7 issues completed
Open

Description

@jeffpaul

Overview

This issue tracks the implementation of embedding support in the WordPress AI plugin. Existing embedding work (#683, #891, #943) should be divided into smaller, independently reviewable PRs rather than combining foundational architecture and user-facing features into large feature PRs.

Proposed Work

Each PR should generally:

  • Branch independently from the develop branch
  • Be reviewable and mergeable on its own
  • Avoid including user-facing features before the required foundation is available
  • Include focused tests and documentation

Related PRs may be stacked when a direct implementation dependency makes that more practical.

Testing

Local models, including Ollama, should be supported as a practical way to test embedding workflows without ongoing provider API costs.

Embeddings generally only need to be regenerated when their source content changes. Similarity calculations and other operations using stored embeddings should not require additional provider requests.

Links to the relevant issues and PRs should be added to the checklist above as the work is defined.

Activity

  1. added this to the 1.4.0 milestone on Aug 20, 2026
  2. added
    [Type] OverviewComprehensive, high level view of an area of focus often with multiple tracking issues
    on Aug 20, 2026
  3. moved this from Triage to In progress in WordPress AI Roadmapon Aug 20, 2026
  4. dkotter commented on Aug 20, 2026

    @dkotter
    Contributor

    Note that the first part of this (pulling in embedding support from the PHP AI Client) was done in #892. After further discussion though, some changes were requested on the PHP AI Client side and we have WordPress/php-ai-client#274 opened now to address that. Once those changes are approved, we'll want to bring those back into the AI plugin and re-enable the loading of this that was removed in #946.

  5. self-assigned this
    on Aug 20, 2026
  6. ColinM-sys commented on Aug 22, 2026

    @ColinM-sys
    Contributor

    Following up from #943 — I tested php-ai-client#274 end-to-end against a local Ollama (full report: WordPress/php-ai-client#274 (comment)). Unit suite green on PHP 7.4–8.5, 36/36 live cases pass, and the no-model auto-selection that #274 removes is reproducible on trunk.

    One concrete item it surfaces for the first checklist entry: #274 removes usingProvider() and usingModelPreference() from EmbeddingBuilder in favour of a required usingProviderModel(). generate_embeddings() in includes/helpers.php calls both removed methods, and the semantic-search callers in #891 / #943 pass model_preference => [ $model ], so re-enabling embedding support once #274 is re-vendored (the #946 follow-up) means moving that wrapper and its callers to an explicit provider + model.

    I'd like to take:

    Both as small PRs off develop, tested against Ollama. Will link them here as they open.

  7. pinned this issue on Aug 26, 2026
  8. 13 remaining items

  9. moved this from Done to In discussion / Needs decision in WordPress AI Roadmapon Sep 16, 2026
  10. modified the milestones: 1.4.0, 1.5.0 on Sep 28, 2026
  11. jeffpaul commented on Oct 28, 2026

    @jeffpaul
    MemberAuthor

    Moving to the 1.6.0 release where we can focus on some of the Experiment PRs that build functionality based on Embeddings.

  12. modified the milestones: 1.5.0, 1.6.0 on Oct 28, 2026
  13. moved this from In discussion / Needs decision to To do in WordPress AI Roadmapon Oct 28, 2026
  14. ColinM-sys commented on Nov 2, 2026

    @ColinM-sys
    Contributor

    Hi @jeffpaul, thanks for the update. For the next embeddings piece I'd like to take retrieval: embedding a query with the consumer's model, matching it against the stored vectors, and returning the closest objects, so the 1.6.0 experiments can all reuse it. I'd keep it to one focused PR off develop. Should I start on it now, or wait for the next version? Or would you rather I take something else first? lmk

  15. dkotter commented on Nov 2, 2026

    @dkotter
    Contributor

    embedding a query with the consumer's model, matching it against the stored vectors, and returning the closest objects

    @ColinM-sys I believe all of that is done or will be done in #1252. At this point, the next steps are building out individual features that utilize embeddings, which we have a few PRs up that do that. May be worth closing this tracking issue out now.

  16. ColinM-sys commented on Nov 2, 2026

    @ColinM-sys
    Contributor

    Oh, I see that now. I didn't check the open PRs for overlap before claiming retrieval. I'll take frontend chat (#142), since it's unassigned. Is that a good one, or would you rather I take something else first?

    Agreed on closing this out. Before you do, the Google embedding PR (#30) and unvendoring the PHP client are still open on the checklist, so they'd need a home. (I think)

  17. dkotter commented on Nov 4, 2026

    @dkotter
    Contributor

    As far as I'm aware, no one is working on #142 so I'm fine if you want to look into that.

    A couple notes I'd keep in mind:

    1. There are likely pieces in Add an "embeddings" mode to Content Classification #1252 that you'd want so need to figure out the best path forward there, whether you wait for that PR to merge before starting, whether it makes more sense for me to pull out any of the more general, reusable pieces from that PR into separate PRs that can be merged quicker or more easily built on top of or something else. Just don't want to duplicate effort here
    2. Something discussed on Frontend chat agent powered by site content #142 but hasn't been decided on is rate limiting. This feature will need to be available to non-logged in users which opens up an easy avenue for abuse. So likely need a decent rate limiting approach to avoid someone spamming things to cause a high bill. That's likely something that could be built in a separate PR first and built in a way that other features can take advantage of it
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

[Type] OverviewComprehensive, high level view of an area of focus often with multiple tracking issues

Projects

Relationships

None yet

Development

No branches or pull requests

Issue actions