Keeping embeddings current

When a record's text changes, its embedding must be made again. Do that in a subscriber, which runs after the change is saved (Subscribers), never in the request that saved it: it is an outside call that can wait.

example

// domains/support/subscribers/embed_article.jadposubscriber embed_article {    on: HelpArticleSaved    run: refresh_meaning(event.article_id)}// domains/support/actions/refresh_meaning.jadpo/// Remakes a help article's embedding from its current text.action refresh_meaning(    article_id: HelpArticle.id) fails EmbeddingDelayed, HelpArticleNotFound -> HelpArticle {    var article: HelpArticle = attempt HelpArticle.get(article_id)    var fresh_meaning: ArticleMeaning = attempt ai.embed(article.body, as: ArticleMeaning) {        timed_out: EmbeddingDelayed        over_budget: EmbeddingDelayed        too_long: EmbeddingDelayed    }    return HelpArticle.update(article) {        set: {            body_meaning: fresh_meaning        }    }}

After you change the embedding model, every record needs a new embedding. Make them in a batch with batches.each (Batches: work over many items), which can resume and stays within the daily limit. An item over budget is recorded as the batch's problem and retried on its next run.

Related