Vercel AI for Bun

Import name: Sentry.vercelAIIntegration

This integration is enabled by default when tracing is enabled. If you'd like to modify your default integrations, read this.

The vercelAIIntegration adds instrumentation for the ai SDK by Vercel to capture agent tracing gen_ai spans, recording model, token usage, latency, and (when enabled) inputs and outputs.

Sentry.init({
  dsn: "___PUBLIC_DSN___",
  // Tracing must be enabled for agent tracing to work
  tracesSampleRate: 1.0,
  integrations: [Sentry.vercelAIIntegration()],
});

Sentry.init({
  dsn: "___PUBLIC_DSN___",
  // Tracing must be enabled for agent tracing to work
  tracesSampleRate: 1.0,
  integrations: [Sentry.vercelAIIntegration()],
});

Supported Operations

By default, tracing support is added to the following ai functions:

• generateText()
• streamText()
• generateObject()
• streamObject()
• embed()
• embedMany()
• rerank()
• experimental_evaluate() — requires Sentry SDK 11.1.0 or higher and AI SDK 7

Plus generate() and stream() on ToolLoopAgent.

Trace Jev Evaluations

With Sentry SDK 11.1.0 or higher and AI SDK 7, vercelAIIntegration automatically traces experimental_evaluate() calls, including calls to TypeSafe's Jev model through Vercel AI Gateway. Keep your existing Sentry setup with tracing enabled; no additional integration or telemetry registration is needed.

Configure AI Gateway authentication with AI_GATEWAY_API_KEY or Vercel OIDC, then call experimental_evaluate():

import { experimental_evaluate } from "ai";

const result = await experimental_evaluate({
  model: "typesafe-ai/jev",
  state: "I was charged twice. Please refund the extra charge.",
  questions: {
    department: {
      type: "choice",
      instructions: "Which team should handle this?",
      criteria: {
        billing: "Charges and refunds",
        technical: "Bugs and outages",
      },
    },
  },
});

console.log(result.answers.department);

import { experimental_evaluate } from "ai";

const result = await experimental_evaluate({
  model: "typesafe-ai/jev",
  state: "I was charged twice. Please refund the extra charge.",
  questions: {
    department: {
      type: "choice",
      instructions: "Which team should handle this?",
      criteria: {
        billing: "Charges and refunds",
        technical: "Bugs and outages",
      },
    },
  },
});

console.log(result.answers.department);

Each call creates a gen_ai.evaluate span with model, latency, and token usage when available. When recording is enabled, Sentry captures state and questions as input messages and answers as output messages. For TypeSafe answers, Sentry also includes confidence values from providerMetadata.typesafe.confidence when present.

Evaluation inputs and outputs follow the same recordInputs and recordOutputs options as other Vercel AI calls. Disable these options if evaluation state or answers contain sensitive data.

The AI SDK's evaluation API is experimental and may change in patch releases. If you use TypeSafeClient.systemOne() directly instead, see the TypeSafe integration.

Supported Versions

• ai: >=3.0.0 <=7

Options

recordInputs

Type: boolean (optional)

Records inputs to ai function calls (such as prompts and messages).

Defaults to true if dataCollection.genAI.inputs is true, which is the default.

recordOutputs

Type: boolean (optional)

Records outputs from ai function calls (such as generated text and responses).

Defaults to true if dataCollection.genAI.outputs is true, which is the default.

Sentry.init({
  dsn: "___PUBLIC_DSN___",
  tracesSampleRate: 1.0,
  integrations: [
    Sentry.vercelAIIntegration({
      recordInputs: false,
      recordOutputs: false,
    }),
  ],
});

Sentry.init({
  dsn: "___PUBLIC_DSN___",
  tracesSampleRate: 1.0,
  integrations: [
    Sentry.vercelAIIntegration({
      recordInputs: false,
      recordOutputs: false,
    }),
  ],
});