Vercel AI SDK

Automatically capture governance traces for every generateText, streamText, generateObject, and streamObject call using the VeriProof Vercel AI SDK instrumentation adapter.

TypeScriptVercel AI SDKNode.js ≥ 18

Prerequisites


npm install @veriproof/sdk-core @veriproof/sdk-instrumentation-ai-sdk ai

You’ll also need an AI provider. This example uses the Anthropic provider:


npm install @ai-sdk/anthropic

Environment


# .env
VERIPROOF_API_KEY=vp_cust_your-tenant.<key-material>
VERIPROOF_APPLICATION_ID=my-ai-app
ANTHROPIC_API_KEY=sk-ant-...

Complete example


import {
  VeriproofClient,
  VeriproofSdkOptions,
  DecisionContext,
  DecisionType,
  RiskLevel,
  SessionIntent,
  SessionOutcome,
} from '@veriproof/sdk-core';
import { VeriproofAISDKInstrumentation } from '@veriproof/sdk-instrumentation-ai-sdk';
import { anthropic } from '@ai-sdk/anthropic';
import { NodeTracerProvider } from '@opentelemetry/sdk-trace-node';
 
// 1. Configure VeriProof once at startup
const provider = new NodeTracerProvider();
const client = new VeriproofClient(
  VeriproofSdkOptions.createDevelopment({
    apiKey: process.env.VERIPROOF_API_KEY!,
    applicationId: process.env.VERIPROOF_APPLICATION_ID!,
  }),
);
const { exporter } = client.configureTracing({
  serviceName: process.env.VERIPROOF_APPLICATION_ID!,
  tracerProvider: provider,
  setGlobal: true,
});
provider.register();
 
// Create the instrumentation adapter — wraps AI SDK calls with OTel spans
const aiSdk = new VeriproofAISDKInstrumentation();
 
// Graceful shutdown — flush any buffered spans before the process exits
process.on('SIGTERM', async () => {
  await exporter.shutdown();
  process.exit(0);
});
 
// 2. Build a governance session
const session = client.startSession(
  'loan-risk-assessment',
  process.env.VERIPROOF_APPLICATION_ID,
  { sessionIntent: SessionIntent.Decision, decisionType: DecisionType.Recommendation },
);
 
// 3. Call the AI framework via the adapter — spans captured automatically
const { text } = await aiSdk.generateText({
  model: anthropic('claude-3-5-haiku-20241022'),
  prompt: 'Assess the creditworthiness of applicant A-1042. Be concise.',
  maxTokens: 256,
});
 
// 4. Record the decision and complete the session
session
  .setDecision(DecisionContext.simple('creditworthiness', text))
  .setRiskLevel(RiskLevel.Medium)
  .setOutcome(SessionOutcome.Success);
 
// 5. Complete the session — this flushes and Merkle-seals the trace
const result = await session.complete();
console.log('Session anchor:', result.merkleRoot);

What you’ll see in VeriProof

After running this example you’ll see a session in the dashboard with:

Span	Key attributes
`session loan-risk-assessment`	`veriproof.session.intent=decision`, `veriproof.decision.type=recommendation`
`ai.generateText`	`ai.model.id=claude-3-5-haiku-20241022`, `ai.usage.inputTokens`, `ai.usage.outputTokens`

The ai.generateText span is captured automatically by VeriproofAISDKInstrumentation — no manual span creation required.

The instrumentation captures model metadata (provider, model ID, token counts) but does not capture prompt or completion content by default. Set enableContentCapture: true in VeriproofSdkOptions to opt in — only do this if your data handling policies permit it.

Streaming

streamText is captured identically — the instrumentation hooks into the stream’s completion event:


import { streamText } from 'ai';
 
const { textStream } = await streamText({
  model: anthropic('claude-3-5-haiku-20241022'),
  prompt: 'Summarise the key risks for this loan application.',
  maxTokens: 512,
});
 
for await (const chunk of textStream) {
  process.stdout.write(chunk);
}
// Span is recorded when the stream is fully consumed

Next steps

Governance Session Builder — add EU AI Act metadata, guardrail results, and human oversight flags
TypeScript SDK Core reference — full VeriproofSession API
AI SDK instrumentation reference — all captured attributes