Add instrumentation for ai

[mydea]

Description

We want to support Vercel AI (https://github.com/vercel/ai). There seems to be some OTEL instrumentation baked in - we need to test this, and ensure this follows the LLM conventions: https://develop.sentry.dev/sdk/telemetry/traces/modules/llm-monitoring/ (which probably means adjusting the emitted spans accordingly).

Tasks

Beta Give feedback

1. Create SDK integration with span transforms
2. Create OTEL instrumentation that injects experimental_telemetry if it is not defined.

[AbhiPrasad]

https://sdk.vercel.ai/docs/ai-sdk-core/telemetry

[AbhiPrasad]

Gave it a try with https://github.com/e2b-dev/ai-artifacts

Looks pretty decent

Currently it collects:

ai.model.id
ai.model.provider
ai.prompt
ai.schema
ai.settings.mode

it doesn't seem to include any information about tokens used or similar, which is problematic for our llm monitoring.

It uses ai.X which matches our conventions more closely, but OTEL conventions seem to use gen_ai.X: https://opentelemetry.io/docs/specs/semconv/gen-ai/gen-ai-spans/

[AbhiPrasad]

Decision: We'll be making this an integration that lives in @sentry/node.

Right now to enable OTEL based telemetry in the SDK, you'll have to add experimental_telemetry: { isEnabled: true } to each individual function call. We'll probably add some OTEL instrumentation that automatically injects this so it doesn't have to be user configured.

[github-actions]

A PR closing this issue has just been released 🚀

This issue was referenced by PR #13892, which was included in the 8.43.0 release.