Vercel AI SDK

Trace every streamText, generateText, and generateObject call.

Install

npm install @trulayer/sdk ai

Instrument

instrumentVercelAI takes the ai module’s functions and returns instrumented replacements. Import it once at the top of any module that uses the Vercel AI SDK.

import { generateText, streamText, generateObject } from "ai";
import { TruLayer, instrumentVercelAI } from "@trulayer/sdk";

const tl = new TruLayer({
  apiKey: process.env.TRULAYER_API_KEY!,
  projectName: "my-app",
});

const ai = instrumentVercelAI(tl, { generateText, streamText, generateObject });

// Use ai.generateText / ai.streamText / ai.generateObject from here on.

Minimal example — streamText in a Next.js Route Handler

app/api/chat/route.ts

import { openai } from "@ai-sdk/openai";
import { TruLayer, instrumentVercelAI } from "@trulayer/sdk";
import { streamText } from "ai";

export const runtime = "nodejs"; // required — edge runtime is not supported

const tl = new TruLayer({ apiKey: process.env.TRULAYER_API_KEY!, projectName: "chat" });
const ai = instrumentVercelAI(tl, { streamText });

export async function POST(req: Request) {
  const { messages } = await req.json();
  const result = await ai.streamText({
    model: openai("gpt-4o-mini"),
    messages,
  });
  return result.toDataStreamResponse();
}

generateObject with a Zod schema

import { anthropic } from "@ai-sdk/anthropic";
import { z } from "zod";

const { object } = await ai.generateObject({
  model: anthropic("claude-sonnet-4-6"),
  schema: z.object({
    sentiment: z.enum(["positive", "neutral", "negative"]),
    reasoning: z.string(),
  }),
  prompt: "Classify: 'The onboarding was great but docs were sparse.'",
});

The span captures the structured object as output and any validation error as span.status = error.

What gets captured

• llm spans around every wrapped call with:
  • input — the last message (or prompt string)
  • output — completed text, or the structured object for generateObject
  • model — the model id (for example gpt-4o-mini)
  • prompt_tokens / completion_tokens
  • latency_ms
• Stream results hold the span open until the final token arrives; usage is recorded from the resolved result.usage promise.

Known gotchas

• Next.js Edge runtime is unsupported. Add export const runtime = "nodejs" to any Route Handler or Server Action that calls an instrumented function. The SDK relies on node:async_hooks for trace context propagation, which is not available on Edge.
• Only wrapped functions are traced. If you import generateText directly from ai elsewhere in your code, those calls bypass TruLayer. Pick one convention per module.
• Tool calls are recorded on the parent llm span. For a dedicated tool span per tool execution, wrap the tool body with tl.span("tool:name", { spanType: "tool" }).