AlloyDB can now extract and classify text straight from SQL

Google added two AI functions to AlloyDB for PostgreSQL on 27 October, both in Preview. ai.extract pulls structured fields out of unstructured text inside a query, defined by a JSON schema or a plain-language description of what you want back, which means a lead-scoring or support-ticket pipeline no longer needs a separate extraction service. ai.classify sorts text into labels you supply, in either single- or multi-label mode. Both default to Gemini 3.5 Flash-Lite and need PostgreSQL 18 or later with the google_ml_integration extension at 1.6.1 or above.

SELECT ai.extract(
  input => bio_text,
  output_schema => '{"email": {"type": "string"}}'
) FROM leads;

The same release also renamed parameters on three existing functions, ai.analyze_sentiment, ai.summarize and ai.agg_summarize: prompt/prompts become input/inputs, and the cursor-based versions drop prompt entirely.

Why it matters: this is SQL absorbing a job that used to mean standing up a separate NLP service and gluing it to the database with a queue. If you’re already on AlloyDB, that’s one less moving part for extraction and classification work.

The caveat: everything here is Preview under Google’s pre-GA terms, and this same release just broke parameter names on functions that shipped months ago. Treat the API as unstable until it’s GA.