Skip to content
Embabel Worlds

7.5 ai.classify — the per-row projection

MATCH (me:AssistantUser)-[:TRACKS]->(n:NewsItem)
RETURN n.title,
       ai.classify(n, 'urgency: high, medium, or low')     AS urgency,
       ai.classify(n, 'topic in one word')                 AS topic

Means: “return the fetched news, and label each item along a dimension there is no column for” — a computed, LLM-decided category rather than a stored field.

How it executes:

  • Fetch as above.
  • Label — one batched call per dimension. Each fetched row’s text is labelled for the named dimension (if the dimension lists categories, the label is one of them; otherwise a short free label). Every row is kept and its label stamped under a per-dimension slug property (ai_classify_urgency…), so two classifications in one RETURN never collide.
  • Rewrite & run. As with ai.score, the executor rewrites each ai.classify(n, '…') in the executed query to coalesce(n.ai_classify_<slug>, '') (blank for any unlabelled row), and the projection returns it.

Use it only for a subjective label no property holds; when a real field already carries the value, return that. (It is a projection, not a filter — to keep only one category, classify and then WHERE label = '…', or use ai.relevant directly.)