On-device query summarization

Get a plain-language preview of a query from a model you control, before you have to parse every clause yourself.

A query you did not write is a small reading exercise before it is anything else. Query summarization turns one into two or three plain sentences: what it does, and what its result actually shows.

What the model is given

The summary is not produced from the SQL text alone. SQLly hands the model the query, the slice of your catalog the query touches, and the facts of the run that just happened:

• the result columns that came back,
• the row count,
• and the elapsed time.

That is why a summary can say “returned no rows, which for this filter means …” instead of describing what the query would do in the abstract. The answer also carries the list of catalog objects it was grounded in, so you can check its sources rather than take its word.

Where it runs

Whichever provider profile you have made active - an on-device model through Ollama or the MLX runtime on Apple Silicon, your own OpenAI- or Anthropic-compatible endpoint, or a cloud provider you deliberately configured. On-device is the honest default: no key, no account, no egress.

In progress: the summary bar, the per-query preference, and the grounded prompt path are built and tested; connecting them to each provider backend is the remaining work.

The same grounding path powers AI review for data-changing queries. For the model plumbing itself, see Intelligence.