AI Data Grid

Documentation

Limitations

What AI Fill doesn't do, and the known limitations of this version.

What AI Fill is not

AI Fill fills cells a developer has configured, when a user or the app asks. It doesn't include:

  • configuration by end users, or a prompt builder;
  • columns that recompute continuously: a result is never re-run by itself, even when it goes stale;
  • learning from corrections;
  • an assisted-dropdown product, semantic lenses, semantic search, conditional formatting or joins;
  • what-if exploration or chat;
  • external enrichment (looking things up outside the row) or generated prose: Jev answers with ids and numbers only;
  • other model providers. AI Fill talks only to Jev.

Answers are model outputs

  • Probability and confidence aren't accuracy. They are what the model reports. Thresholds are your choice; evaluate them on your own data before relying on them, and before turning on autoApply.
  • Answers can change between model versions. jev-latest is an alias. Every answer keeps the versioned model id that produced it (for example jev-1.13.0), so you can tell which model answered.

Browsers can't call Jev directly

TypeSafe's API currently rejects browser CORS preflights from every origin tried, so direct mode works only from Node. In a browser, it refuses without dangerouslyAllowBrowser, and with it the call fails as a network error that points to endpoint mode. Browser apps need endpoint mode, for example a route built with createJevHandler; for demos, use scripts/jev-dev-proxy.mjs. See Connecting to Jev.

Filtered-out rows

AI Fill knows a row only through getRowId and getRowIndex on the rows that are displayed, so it can't tell a row hidden by a filter from a deleted one. Wherever it has to find a row, a filtered-out row counts as deleted, and its result is dropped with reason: "row-missing" (reported through onResult) and nothing is written:

  • an answer that arrives while its row is filtered out. Fill the row again once it's displayed;
  • notifyRowsChanged() called while a filter hides the row (with no ids, or with its id): its decided results are dropped;
  • accept({ cells }) naming a hidden row, and an accept or Choose from the UI that reaches one, for example an inspector left open while a filter hides its row.

Otherwise a decided result on a hidden row keeps its record, "Accept all eligible" skips it, and rejecting or re-running never drops it. Storybook story 10 Sorting and filtering while pending shows both sides.

Persistence and undo

  • Suggestions live in memory. They aren't saved, and a reload starts with none. Only accepted values persist, through your own edit handlers. Commit ids, and so revertCommit, last as long as the grid's AI Fill session.
  • useUndoRedo is position-based. An undo after re-sorting or filtering writes to display positions, not to the rows that were accepted. api.revertCommit(commitId) is the id-safe alternative.
  • An onCellsEdited that returns true hides AI Fill's writes from useUndoRedo, as it does for paste, because the per-cell onCellEdited calls don't happen.

Requests

  • One row per request. Rows are never packed together, so a fill over 500 rows sends at least one request per row (fewer when answers are cached): one for each distinct row state and model, with at most maxQuestionsPerRequest questions each. Two AI columns with different sources over 500 rows send up to 1,000 requests, within concurrency and maxRequestsPerMinute.
  • Jev's error-body shape isn't fully documented, so AI Fill reads the message from the fields it knows. An unexpected shape still maps to the right error kind by its HTTP status, with a generic message.

Accessibility

The grid's canvas doesn't expose suggestions to screen readers. The menus, the confirm dialog, the inspector and the status bar are accessible DOM, and a live region announces what menu, inspector and shortcut actions did. Screen-reader users review results through the inspector.

API

  • AIFillRun (the return type of fill, retry and rerunStale) and AIActiveRun (an element of getRunState().active) aren't exported by name. Use inference, or ReturnType<AIFillApi["fill"]>.
  • require("@specstory/ai-data-grid/server") needs a Node version that can require ES modules: 20.19+, 22.12+ or 24. Older versions must use import.