Just saw the beta flag in the dashboard. The feature claims to analyze your video transcript and suggest stock B-roll clips to insert automatically.
Has anyone with CLI/API access tested the logic behind the suggestions? I'm skeptical about how it maps keywords to relevant clips without manual tagging. Opus is usually good at the audio/text side, but this seems like a leap into a different problem space.
Initial questions:
* What's the source of the B-roll library? Licensed or requiring upload?
* Can you define your own clip library via API?
* How are suggestions weighted? Pure keyword match or semantic analysis?
If it's just a basic keyword scrape, it'll be useless for technical content. Need to see the output schema before integrating into any pipeline.
-c
The library source is their internal stock collection, licensed but no third-party uploads yet. However, you can define a custom library via the API, though it's currently undocumented. You need to use the `projects/{id}/assets` endpoint with a `type=video` query and a POST with a signed URL. I've got a snippet that works.
On your weighting question, it's not just keyword matching. They're using their existing transcript embeddings and running a cosine similarity against clip descriptions, which explains why it's still in beta. For technical content, it'll likely fail until you can feed it your own pre-tagged asset library. The output schema includes a confidence score and the matched keyword phrase, but not the underlying vector.
I'd hold off on pipeline integration until they expose the embedding model parameters. Without that, you can't tune for specificity.
IntegrationWizard