AI Agent Examples

AI Metadata quickstart

Inngest can automatically extract AI metadata — model, input/output tokens, provider, latency, and cost when available — from OpenTelemetry spans created while a step runs, and attach it to that step as inngest.ai metadata. It's on by default; you don't need to add middleware or call anything from your function code.

This guide covers the minimum setup: getting an OpenTelemetry provider loaded, and making sure your AI calls actually emit the spans Inngest reads. For the full picture — bring-your-own provider, CJS/ESM preload details, and Extended Traces — see Set up OpenTelemetry with Inngest.

AI metadata extraction is independent of Extended Traces — you don't need extendedTracesMiddleware() for this to work.

Step 1: Load an OpenTelemetry provider

AI metadata extraction reads attributes off OpenTelemetry spans, so a provider needs to be registered in your process before it starts. If your app doesn't already set up OpenTelemetry, the fastest path is @inngest/otel:

npm install @inngest/otel

node --import @inngest/otel/node ./app.js

If your app already has its own OpenTelemetry provider, keep it — see manual setup for the CJS/ESM preload details.

Step 2: Make sure your AI calls emit gen_ai.* spans

Inngest's extractor only reads spans with OpenTelemetry GenAI semantic convention attributes (gen_ai.request.model, gen_ai.usage.input_tokens, and so on). Depending on how you call your model, you may need to do one of the following:

  • OpenAI, Anthropic, or Google Generative AI SDKs directly: @inngest/otel's preload auto-instruments these SDKs and emits gen_ai.* spans for you — nothing else to configure once Step 1's preload is loaded.
  • Vercel AI SDK: the ai package emits gen_ai.* attributes natively, but telemetry is opt-in per call. Set experimental_telemetry: { isEnabled: true } on each generateText/streamText/etc. call you want captured:
import { generateText } from "ai";

await generateText({
  model: openai("gpt-4o"),
  prompt: "...",
  experimental_telemetry: { isEnabled: true },
});

If you're using a different SDK or a custom HTTP client, it needs to emit spans with the same gen_ai.* attributes for Inngest to pick them up — check the OpenTelemetry GenAI conventions for the attribute names to set.

Step 3: Verify it in a run

Trigger a function that makes one of the calls from Step 2, then open that run's trace in the Inngest dashboard. The step should show an AI Metadata entry with the model, token counts, and latency for that call.

If you don't see it:

  • Confirm the OpenTelemetry preload from Step 1 is loading before your app code (check for a startup log line, or see the ESM dual-package hazard section if spans go missing silently).
  • Confirm the call in Step 2 is actually instrumented — for Vercel AI SDK, double check experimental_telemetry.isEnabled is set on that specific call.
  • Confirm aiMetadata hasn't been set to false on your Inngest client. It defaults to on (aiMetadata: true):
import { Inngest } from "inngest";

export const inngest = new Inngest({
  id: "my-app",
  aiMetadata: false, // omit this line to keep AI metadata extraction on
});

More context