Placing media in the request
You never encode media yourself. Reference media.frames where the model
expects the encoded bytes, and Elemental Inference substitutes
them after your expression has been evaluated. While the expression runs, those values are
placeholders, so the media itself is never exposed to the expression.
The consequence is one rule: place a media value whole.
-
Assign an item of
media.framesdirectly to a field. Don't concatenate it, take a substring of it, or otherwise transform it. -
Don't place the same frame in two places in the same request. It doubles the size of the request and what the model charges for it.
A request expression that breaks either rule is rejected when you save the function, or, in a branch that the sample context doesn't reach, fails at run time with a guardrail error. For more information, see Expression limits and guardrails.
media.transcript, media.contextual, and media.clip
are ordinary data, not media payloads. You can test them, concatenate them, iterate over
them, and reshape them freely.
Use the . map operator to expand media.frames into the array
shape that the model expects, and $append() to place that array after your
prompt text. Inside the map, $ refers to the current frame.
$append( [{ "type": "text", "text": "Describe these frames." }], media.frames.{ "type": "image", "source": { "type": "base64", "media_type": "image/jpeg", "data": $ } } )