View a markdown version of this page

Creating a function - Elemental Inference

Creating a function

A function is an account-level resource. Creating one has no effect on any feed and starts no processing – you attach it to a feed in a separate step, described in Attaching a function to a feed.

A function has the following properties.

Property Required Description
name Yes 1–128 characters. Must begin and end with a letter or a number, and can contain letters, numbers, hyphens, and underscores.
target Yes Either bedrock (invoke an Amazon Bedrock model) or passthrough (invoke nothing).
requestExpression Yes JSONata that builds the request. 1–5,120 characters. See Writing function expressions.
responseExpression No JSONata that selects what to keep from the result. 1–5,120 characters. If you omit it, Elemental Inference stores the whole result.
description No Up to 1,024 characters. Can contain letters, numbers, spaces, underscores, and the characters - . , ' @ : ;
tags No Tags for the function. You can set them only when you create the function; after that, use TagResource and UntagResource.

For a bedrock target you also specify the following.

  • api – the Amazon Bedrock API to call. INVOKE_MODEL is the only value.

  • model – a model ID, an inference profile ID, or the ARN of a custom or provisioned model. Elemental Inference invokes the model in the feed's Region. To use a model in other Regions, specify a cross-Region inference profile ID. A custom or provisioned model must be in the feed's Region.

The response to a create or get request also includes fields that Elemental Inference computes and that you can't set.

  • id and arn – the identifier of the function. You use the ARN when you attach the function to a feed.

  • status – see Function status.

  • references – the IDs of the feeds that currently use this function.

  • contextTypes – the media context that Elemental Inference derived from your request expression. See What a function can read.

When you create or update a function, Elemental Inference rejects it with a ValidationException unless the following are true.

  • Both expressions are valid JSONata.

  • The request expression reads at most one of media.transcript, media.contextual, and media.clip. It can combine that one with media.frames and media.video.

  • A passthrough request expression doesn't read media.frames or media.video. There is no model to send them to.

  • The request expression runs without error against sample media context. For a bedrock target, it must produce a JSON object, because it becomes the request body. For a passthrough target, the request expression, followed by the response expression if there is one, must produce a value. A bedrock response expression is checked for syntax only, because the model response isn't known until the function runs.

Topics