Teer Serverless

@teerai/serverless provides an OTEL compliant tracer for the AI-SDK. It observes ai operations and forwards them to Teer for tracking and analytics. It is optimized for edge and serverless environments.

Quick Start

import { TeerEdge } from '@teerai/serverless'
import { generateText } from 'ai'

const anthropic = createAnthropic({
  apiKey: {{ANTHROPIC_API_KEY}},
})

await generateText({
  model: anthropic('claude-3-haiku-20240307'),
  prompt: 'What is the capital of Ireland?',
  experimental_telemetry: {
    isEnabled: true,
    tracer: TeerEdge.getTracer({ apiKey: {{TEER_SECRET_API_KEY}} }),
    // Unique ID for operation
    functionId: 'test',
    // Opt out / in to tracking the inputs and outputs of your AI-SDK traces
    recordInputs: false,
    recordOutputs: false,
  },
})

// Ensure telemetry is sent before request ends
waitUntil(TeerEdge.forceFlush())

Configuration Options

| Option | Type | Required | Default | Description | | --------------- | --------------------------------- | -------- | ----------------------- | ---------------------------------------------------- | | apiKey | string | true | - | Authentication key for Teer backend | | baseURL | string | false | https://track.teer.ai | Custom endpoint for telemetry data | | debug | boolean | false | false | Enables detailed debug logging | | flushInterval | number | false | 5000 | Milliseconds between automatic flush attempts | | customFetch | typeof fetch | false | fetch | Custom fetch implementation for special environments | | onExport | (spans: ReadableSpan[]) => void | false | undefined | Hook called before spans are exported |

Flush vs Shutdown

The library provides two methods for ensuring telemetry data is sent:

• TeerEdge.forceFlush(): Immediately sends pending spans while keeping the tracer active. Use this when you want to ensure data is sent but expect more telemetry in the future.
• TeerEdge.shutdown(): Sends pending spans and closes the tracer. Use this at the end of your application lifecycle when no more telemetry will be generated.

For serverless environments, forceFlush() is typically sufficient as the runtime will handle cleanup.

Best Practices

Flush

• Set appropriate flushInterval based on your workload:
  • Lower values (1-2s) for burst workloads
  • Higher values (5-10s) for steady workloads

Logging

• Enable debug logging during development:

TeerEdge.getTracer({
  apiKey: 'your-api-key',
  debug: true,
})

waitUntil

• Use waitUntil in edge functions to ensure telemetry is sent:

Vercel

• Docs

export function GET(request: Request) {
  // ... your code ...
  waitUntil(TeerEdge.forceFlush())
  return new Response(`Hello from ${request.url}, I'm a Vercel Function!`)
}

Cloudflare Workers

• Docs

export default {
  async fetch(request, env, ctx) {
    // ... your code ...
    ctx.waitUntil(TeerEdge.forceFlush())

    return res
  },
}