Skip to content

wrapAISDK copies AI SDK v7 file bytes one key per byte instead of turning them into attachments #2750

Open

Description

What happens

With ai 7.0.145 and braintrust 3.32.0, file parts reach the doStream model span as { type: 'file', mediaType, data: { type: 'data', data: Uint8Array } }, the AI SDK 7 shape (SharedV4FileData). processContentPart passes part.data, the wrapper object, to convertDataToAttachment, which only handles a string, Uint8Array, Buffer or Blob, so no attachment is made. deepCopyEvent then runs JSON.stringify and JSON.parse on the raw Uint8Array, which becomes an object with one key per byte.

https://github.com/braintrustdata/braintrust-sdk-javascript/blob/braintrust%403.32.0/js/src/instrumentation/plugins/ai-sdk-plugin.ts#L1297-L1301

The same code is in 3.43.0 and on main as of e0b7107.

Impact

1 MB of image bytes becomes 1,000,000 keys and about 120 MB of heap on every model call. We call Claude through Vertex, which takes no image URLs, so the AI SDK downloads every image in the conversation and each call carries all of them. A chat with 25 MB of images pushed the heap to about 3.1 GB, and our 4 GB serverless instances were killed for memory.

On a single 38-image call (24.5 MB of images), the heap peaked at 2,651 MB with wrapAISDK and at 117 MB without it, and deepCopyEvent alone allocated 5.4 GB.

Reproduction

processContentPart leaves this shape untouched, so the model span event is copied like this.

import { deepCopyEvent } from 'braintrust';

const copy = deepCopyEvent({
  input: [{ type: 'file', mediaType: 'image/png', data: { type: 'data', data: new Uint8Array(1_000_000) } }],
});
// copy.input[0].data.data is a plain object with 1,000,000 keys, about 11 MB of JSON

Workaround we use

A one-line change in processContentPart, applied with pnpm patch to every build in dist, fixes it for us.

       const fileAttachment = convertDataToAttachment(
-        part.data,
+        part.data.type === "data" ? part.data.data : part.data,
         part.mimeType || part.mediaType,
         part.name || part.filename
       );

With it, the same 38-image call peaked at 130 MB of heap with tracing on. A complete fix could also handle { type: 'url' } and the other SharedV4FileData variants in convertDataToAttachment.

Environment

  • braintrust 3.32.0, also seen at runtime on 3.42.0 and in the source of 3.43.0
  • ai 7.0.145 with @ai-sdk/google-vertex 5.0.117 (Anthropic on Vertex)
  • Node 22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions