Vision and media understanding

Take a media input (an image, a video, an audio clip, a document) and produce text about it: a direct answer, a caption, tags, extracted fields, a transcript, a summary. The lever is matching the input modality to a model that accepts it, then asking for exactly the text shape you want.

Inputs to collect

• The media to understand, as a URL or base64 (image, video, audio, or a document rendered to image/PDF).
• The input modality - image, video, or audio. This is the primary routing decision.
• What text you want back: a free-form answer, a one-line caption, a tag list, structured fields (JSON), a transcript, or a summary.
• Optional: the question or instruction (visual Q&A), and whether output must be machine-parseable (then ask for JSON explicitly).

Models

Route by input modality. Confirm the live model + its schema via the runware-models + runware-run skills before calling - never hardcode a stale choice.