What happens quietly after you upload

Enrichment is the queue of background jobs that runs after every upload: writing a one-sentence description for every column, detecting and encoding categorical columns into a form models can use, building the vector and full-text search index chat's retrieval relies on, generating a plain-language dataset summary, and scanning for notable relationships, label problems and anomalies. It also precomputes Factor Finder's ranking, Segments' default grouping, and Regression's default fit, so those pages open on a rendered result instead of a spinner.

The same on-device stack also runs sentiment analysis, keyphrase extraction and topic modelling over free-text columns, chat-answerable rather than kept in a separate tool.

• The deep profile report is on-demand only: it is not part of the automatic queue, you generate it explicitly from the Profiler.
• If the local model fails while ordering a categorical column (lowest to highest), CLAIR does not guess an alphabetical order. It falls back to an unordered mapping and marks it as needing your review, rather than risk a wrong order feeding straight into Factor Finder or a trained model unreviewed.
• A long enrichment call cannot block a chat question indefinitely: chat gives it up to about 12 seconds before moving on to another way of answering.