Optimizing Catalog Metadata with AI

AI Optimization analyzes the active catalog and creates confidence-scored recommendations for supported descriptive metadata such as mood, genre, energy, keywords, instrumentation, and sync use cases. CatalogIQOS generates suggestions for human review. It does not automatically overwrite canonical metadata. A field changes only when an authorized user chooses Apply Suggestion or confirms Apply Selected.

The page guides one cycle through Analyze → Review → Apply → Results. Its recommended action responds to the latest Optimization Run and its pending, applied, and rejected suggestions.

1. Choose the active catalog and run AI Optimization.
2. Wait for the background Optimization Run to complete.
3. Compare Current Value and Suggested Value for each recommendation.
4. Apply supported changes or reject recommendations you do not want to use.
5. Review Fields Improved, Optimization History, and Optimization Runs.

Run it after a spreadsheet import, after Audio Intelligence has added analysis context, after manual metadata cleanup, before Catalog Audit or Export, or after major catalog changes. Finish reviewing pending suggestions before rerunning. Starting a new cycle with unresolved work is blocked because a new run supersedes the previous pending set.

The Workspace Context bar shows the active Workspace and Catalog. CatalogSwitcher reloads AI Optimization for another authorized catalog. Runs, scores, suggestions, and history are catalog-specific, so confirm context before applying metadata.

When the guide says Ready to optimize this catalog, use Run AI Optimization. CatalogIQOS queues an AI_OPTIMIZATION processing job and creates an Optimization Run. While it is queued, processing, or retrying, the guide displays Optimization in progress and links to Processing Queue at /jobs. You may work elsewhere while it completes. Duplicate runs are blocked while one is active.

The job analyzes saved catalog and enrichment signals and creates recommendations; it does not apply them. A failed run shows a safe failure state and links to processing details. Provider stack traces are not displayed in the workflow.

When pending suggestions exist, Review Suggestions opens #suggestions. Each row shows Track, Field, Current Value, Suggested Value, Confidence, review status, and the Optimization Run as source. Confidence represents the strength of the implemented inference, not verification or authoritative evidence.

Use status, confidence, and field filters to create a manageable set. Apply Suggestion writes the supported value to canonical Track or Track Enrichment metadata and marks the suggestion accepted. Reject Suggestion means “do not apply this suggestion”; it does not declare the source data permanently invalid. The default Pending filter makes a reviewed suggestion disappear from the queue. Select Accepted, Rejected, or All statuses to find it again.

For individual recommendations, compare against an authoritative source and choose Apply or Reject. For a homogeneous low-risk group, use Select All, Select All Filtered, and Apply Selected. Before bulk application, CatalogIQOS confirms the number of changes and warns that canonical metadata will be updated. Bulk actions and individual mutations are disabled for read-only users and Enterprise Demo viewers.

Applied mood and genre suggestions update Track fields. Supported mood, energy, instrumentation, keywords, and sync-use-case suggestions update Track Enrichment; keywords also connect Track tags. CatalogIQOS records an accepted suggestion and an audit event. Applying is not a claim that the suggested value was independently verified.

Fields Improved counts accepted suggestions only; rejected decisions are not improvements. Optimization History distinguishes Applied and Rejected rows and shows Track, field, previous value, reviewed value, and source run. Optimization Runs shows recent start time, tracks processed, generated-suggestion count, and status.

The current model does not store a dedicated applied timestamp or actor directly on each suggestion. Actor and application details are retained in audit logs, but the page cannot yet show them as first-class History columns. A trustworthy before/after Optimization Score snapshot is also not stored per run, so the guide does not claim a score delta.

Rerun after finishing the current review and after meaningful metadata or analysis changes. In the Complete state, use Run AI Optimization Again. With pending suggestions, the page prioritizes Review and both the interface and API block a new run. An active queued/processing/retrying run also blocks duplicates.

### Does AI change my metadata automatically? No. Analysis generates pending suggestions. Only Apply actions change supported canonical fields.

### What does confidence mean? It is an inference-strength signal used for prioritization, not proof or verification.

### Why do I have no suggestions? The catalog may not have been analyzed, supported fields may already contain usable data, or available signals may not produce a new recommendation. Check the latest run status and filters.

### Can I reject a suggestion? Yes. Reject records a review decision and leaves canonical metadata unchanged.

### Can I undo an applied suggestion? There is no one-click undo in AI Optimization. Use Optimization History to identify the previous value, then review and edit the canonical Track record deliberately.

### Why is my Optimization Score still low? The score combines current enrichment signals and may reflect unresolved catalog issues beyond the applied field. Apply relevant suggestions, review the Track, and rerun analysis or Catalog Audit when appropriate.

### Why are suggestions still pending? Generating a run never applies recommendations. An authorized reviewer must Apply or Reject each pending item, individually or in a selected batch.

• Run AI Optimization disabled: verify an editable role and a non-demo workspace. Finish pending review and wait for any active run to reach a final state.
• Optimization remains processing: open View Processing Queue and inspect the AI Optimization job. Do not submit duplicates while queued, processing, or retrying.
• No suggestions generated: confirm the run completed, clear restrictive filters, and review whether supported fields already contain values or lack enough analysis context.
• Bulk Apply disabled: select pending rows and confirm your role permits metadata changes. Accepted/rejected rows and Demo Viewer records are not selectable.
• Suggestion disappears: reviewed items leave the default Pending filter. Choose Accepted, Rejected, or All statuses.
• Applied change not reflected: allow the request and page refresh to complete, clear filters, then open the canonical Track. If it still differs, check the error banner and avoid applying twice.
• Failed optimization run: use View Processing Details, then Retry Optimization after the job is no longer active and no pending review remains.
• Score did not change: a single applied field may not change the aggregate score, and the model does not store per-run before/after snapshots. Review remaining gaps and rerun only after completing the cycle.

AI Optimization assists catalog review. Rights, ownership, identifiers, clearance, and contractual facts require authoritative sources and human approval.