v5-mirror: serialize UInt8Arrays as base64 for inner telemetry spans

[dinmukhamedm]

Background

Clone of #6357 but for v5.

Description below is copy-pasted from there

generateObject, generateText, streamText, and streamObject currently call JSON.stringify on the input messages. If the input messages contain an image, it is most likely normalized into a Uint8Array.

JSON.stringify does not the most obvious things to TypedArrays including Uint8Array.

// this returns '{"0": 1,"1": 2,"2": 3}', where I'd expect this to be '[1,2,3]'
JSON.stringify(new Uint8array([1, 2, 3]))

In practice, this results in bloating images by about 5-15x depending on the original image size. For Laminar, for example, a span with 3 avg sized images will not be able to be sent as it is larger than the (reasonably high) gRPC payload size for our traces endpoint.

From MDN docs:

// TypedArray
JSON.stringify([new Int8Array([1]), new Int16Array([1]), new Int32Array([1])]);
// '[{"0":1},{"0":1},{"0":1}]'
JSON.stringify([
  new Uint8Array([1]),
  new Uint8ClampedArray([1]),
  new Uint16Array([1]),
  new Uint32Array([1]),
]);
// '[{"0":1},{"0":1},{"0":1},{"0":1}]'
JSON.stringify([new Float32Array([1]), new Float64Array([1])]);
// '[{"0":1},{"0":1}]'

Summary

Added a function that maps over messages in a LanguageModelV2Prompt and maps over content parts in each message, replacing Uint8Arrays with raw base64 strings instead.

Call this function when calling recordSpan for the inner (doStream/doGenerate) span in generateObject, generateText, streamText, and streamObject.

Verification

Ran this small script against a local instance of Laminar and logged the Telemetry payloads (span attributes) on the backend to verify that they are indeed base64.

import { Laminar, getTracer } from '@lmnr-ai/lmnr'

Laminar.initialize();

import { openai } from '@ai-sdk/openai'
import { generateText, generateObject, streamText, streamObject, tool } from "ai";
import { z } from "zod";
import dotenv from "dotenv";

dotenv.config();

const handle = async () => {
  const imageUrl = "https://upload.wikimedia.org/wikipedia/commons/b/bc/CoinEx.png"
  const imageData = await fetch(imageUrl)
    .then(response => response.arrayBuffer())
    .then(buffer => Buffer.from(buffer).toString('base64'));

  const o = streamObject({
    schema: z.object({
      text: z.string(),
      companyName: z.string().optional().nullable(),
    }),
    messages: [
      {
        role: "user",
        content: [
          {
            type: "text",
            text: "Describe this image briefly"
          },
          {
            type: "image",
            image: imageData,
            mimeType: "image/png"
          }
        ]
      }
    ],
    model: openai("gpt-4.1-nano"),
    experimental_telemetry: {
      isEnabled: true,
      tracer: getTracer()
    }
  });

  for await (const chunk of o.fullStream) {
    console.log(chunk);
  }
  await Laminar.shutdown();
};

handle().then((r) => {
    console.log(r);
});

Tasks

• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
  • telemetry is experimental, so I reckon a doc update for this small fix is not required
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
  • The destination is not main, and v5 is now in alpha releases, so I wasn't sure of the procedure here
• Formatting issues have been fixed (run pnpm prettier-fix in the project root)

Future Work

Related Issues

Fixes #6210 for v5

[lgrammel]

please add a patch changeset similar to main