The service selects a small Learn context

When a Moow AI chat request is prepared, the API loads published Learn posts from the blog repository, considering up to 200 records. It scores word overlap between the current question and each story’s title, tags, takeaways, excerpt, and section paragraphs. Title matches receive the strongest weight in that code path.

Stories with a positive match are ranked, and no more than four are supplied to the answer process. In Learn mode, if no word match is found, the service falls back to available ranked stories rather than silently inventing a match. This is straightforward lexical retrieval, not a claim that the four selections are always semantically perfect.

The supplied article content is bounded and labeled

For each selected story, the context includes its citation slug, title, category, reviewed date, excerpt, takeaways, and a bounded selection from up to four sections. Paragraph text is length-limited before entering the context. This keeps the grounding set compact and gives the model recognizable editorial fields rather than an unlimited blog export.

The prompt marks the server context as data, not instructions, and tells the model not to follow commands found inside article or user-record text. In Learn mode, the context builder does not add personal profile, meal, progress, workout, or movement records, and the answer request does not include recent conversation history.

Source slugs are checked before display

The answer instructions say to include a story’s exact citation slug when relying on it and to return no source slug when no supplied story supports the response. After generation, the server compares every returned slug with the allowed sources in that request, removes unknown values, and deduplicates repeated selections.

The resulting source metadata contains the slug, title, category, and reviewed date. In the mobile interface, a source card shows the title and category and, when present, the reviewed date. Opening the card starts a Learn reader that requests the full published story by its slug, so the user can inspect more than the answer excerpt.

Understand what this design does not prove

A displayed source shows which supplied Learn story the answer used; it does not guarantee that every sentence is correct, that retrieval found every relevant story, or that a reviewed date makes information timeless. The current selection is word-overlap based, and only a bounded portion of each chosen article is placed in context.

Moow AI’s prompt also requires cautious general health guidance and referral when warranted, but it does not replace a clinician or emergency service. Read the linked story, check its references, and use professional care for personal diagnosis or treatment. Source transparency is a tool for scrutiny, not a reason to suspend it.

Sources

1. NIST: Privacy Framework
2. FTC: Mobile health app privacy and security
3. HHS Surgeon General: Health misinformation