Skip to main content
info

Access Athinia Catalyst at catalyst.athinia.io.

Study data answers

When you open Assist from within a study, you can ask about that study's data in plain language — "average overlay error by tool?", "how many wafers failed process control?", "distribution of risk score by lot" — and Assist queries the dataset to answer.

Study data answer
A data answer: the query result shown as a markdown table with a brief takeaway.

When it's available​

  • Inside a study. Outside a study, Assist still answers platform questions but has no dataset to query.
  • With a running worker. Queries run on the study's compute. If the worker isn't active, Assist says to start it, and answers only platform questions until then.

How it works​

  1. Your data becomes a queryable table. When Assist opens in a study, the dataset is exposed to it as a table named dataset, with its column list, so Assist knows what it can ask about.
  2. Your question is translated to a query. Assist writes a single read-only SQL query (Polars SQL dialect) over dataset.
  3. The query runs on your study's worker, isolated so it can only read your study's dataset and cannot modify anything.
  4. The answer is computed from the result. Assist states the result and the conditions it covers, with at most one brief takeaway from the returned rows. Multi-row results are shown as a table.
  5. It self-corrects. If a query errors, Assist reads the error and retries with a corrected one.

What it respects​

The columns Assist can query come from your study's column configuration, not from the chat:

  • Visibility: hidden columns are excluded; Assist can't query or return them.
  • Display names: columns use their configured names, so you can ask using the terms you see in the study.
  • Data types: columns use their configured types, so numeric, categorical, and date operations behave correctly.

What you can ask​

  • Aggregates and breakdowns: "average yield by site", "count of failures by category".
  • Filtered questions: "how many wafers exceeded the spec limit?".
  • Dates and durations: group by year, compare timestamps, measure time between events ("average days between start and completion?").

Notes and limits​

  • Data answers come only from your dataset. Assist never invents or estimates figures; if the data doesn't answer your question, it says what the data does and doesn't show.
  • Results are capped. Large results are truncated to a sample for display, and Assist flags when this happens. Ask a more aggregated question (counts, averages, top-N) for the full picture.
  • AI responses may not always be accurate; verify important results.