Use Assistant Engine with human review

Assistant Engine is an optional workspace for caption drafts, number/roster review, search, and explainable image ranking. Configure an available provider only if you intend to use it, and understand where it processes images. A model result is a suggestion, not verified identity or editorial truth.

Prepare and request assistance

1. Enable the included Assistant Engine in Plugins. Open Settings → AI Services and inspect the provider/capability configuration your build supports.
2. Select a small group of local images relevant to one task. Start with a non-sensitive sample when testing a provider.
3. Open the Assistant Engine task for caption drafting, number indexing, search, or ranking. Read the reported provider availability and any fallback notice.
4. Run the requested analysis and wait for its result. Compare caption facts to the photograph and reliable event information.
5. Correct uncertain identities, numbers, or wording before approving results. Use the normal metadata/save action only for reviewed text.

Review athlete numbers and roster matches

Import verified people/team/roster data through Data Manager. In the Athlete Event workflow, set the event context, index candidate bib/jersey numbers, and inspect match rows. Correct athlete/team/group values, leave ambiguous matches unapproved, and create athlete-specific Selection Sets only from the rows you have approved. Export/reimport the roster-match CSV when a human correction round is needed; inspect filename/path matches again after import.

Search and ranking boundaries

Local filename/metadata semantic search is different from image-content search using an optional embedding provider. If the provider is absent, read the fallback instead of interpreting text matches as visual recognition. Explainable quality ranking uses local signals such as sharpness, exposure, clipping, entropy, and resolution. It does not establish peak action, expression, story value, or which photograph should be published.

Face recognition depends on a compatible available provider and explicit review. A retained provider contract does not mean a bundled face model works in your build. Do not organize or label people on unreviewed matches.

Verify and troubleshoot

Compare results with known sample images before using a batch. Reopen created sets and any saved captions. If an action is unavailable, inspect provider configuration and plugin state. Do not silently substitute identity guesses for failed analysis. The initial core Burst workflow does not require AI. For teaching interpretation use Camera Decision Coach.