Trainings

Training computes an embedding (a numeric representation of meaning) for every element in the list — that is what makes semantic search possible. Editing elements never launches anything by itself: it leaves one pending training and you decide when to launch it.

Every list has at most one pending training. When you launch it, it moves through these states:

• Pending

  The list changed since the last training. Waiting for you to launch it.

• Queued

  Queued: it starts as soon as there is a free slot (trainings are dispatched fairly across users).

• Training

  The worker is computing embeddings and enriching elements. Usually a few minutes.

• Completed

  Done: the training becomes "In use" and searches use its model immediately.

• Failed

  Something went wrong. The error is shown on the training and you can launch again.

Each launch uses the embedding model you pick. Models trade quality for speed: larger ones understand nuance better but take longer to train and answer slightly slower.

You can train the same list with several models and compare them from the Search page before deciding which one stays in use.

The training marked "In use" is the one answering searches for the list. Completing a training puts it in use automatically.

You can switch back to a previous completed training with "Use" — as long as it covers exactly the elements the list has now. If elements changed since, retrain instead.

• One run per list: a list with a training in progress cannot launch another until it finishes or fails.

• Platform-wide cap: each run reserves real compute, so the number of simultaneous trainings is limited. If a launch is rejected, retry when a slot frees up.