Text remains the primary retrieval channel for every modality. Images and frames confirm and locate; they rarely lead, and a design that leads with them retrieves confidently and wrongly.
Every modality names its known limit on the page. A modality with no stated limit has not been measured, which is more dangerous than one that is simply weak.
Citations for media anchor to a location a person can open — a page and bounding box, a slide number, a player timecode.
Numbers
6% of queries carry media; field-engineer photo search is 78% of those.
Multimodal adds 400 ms and 0.003 USD to a query; the ingest-time cost is far larger and is paid once.
Transcript-led retrieval on video reached 0.81 Recall@20 against 0.44 for keyframe-led, which is why the order is what it is.
Risks
Dense engineering schematics under-retrieve. The caption and surrounding text carry the recall, so a diagram in a document with no caption is effectively invisible — reported to stewards as a coverage defect.
Speaker attribution from audio is weak and is never used as an authority signal.
