LlamaIndex.TS

LlamaIndex.TS is the TypeScript edition of LlamaIndex. Its documented observability path is OpenLLMetry / Traceloop, which is built on OpenTelemetry and emits standard OTLP — so you point its exporter at Logfire’s OTLP endpoint with your write token.

Installation

Terminal

npm install @traceloop/node-server-sdk llamaindex @llamaindex/openai @llamaindex/workflow zod

Usage

Initialize OpenLLMetry before dynamically importing LlamaIndex. This lets OpenLLMetry install its module hooks before the framework loads:

main.ts

import * as traceloop from '@traceloop/node-server-sdk';
import { z } from 'zod';

traceloop.initialize({
  appName: 'llamaindex-ts-logfire',
  disableBatch: true, // flush immediately (dev only)
  baseUrl: 'https://logfire-us.pydantic.dev', // EU: https://logfire-eu.pydantic.dev
  // Use `headers` (raw token), NOT `apiKey` — `apiKey` would send "Authorization: Bearer <token>",
  // which Logfire rejects.
  headers: { Authorization: process.env.LOGFIRE_TOKEN ?? '' },
});

async function main() {
  const [LlamaIndex, { openai }, { agent }] = await Promise.all([
    import('llamaindex'),
    import('@llamaindex/openai'),
    import('@llamaindex/workflow'),
  ]);

  let toolCalls = 0;
  const lookupIncident = LlamaIndex.tool({
    name: 'lookup_incident',
    description: 'Look up an incident by ID.',
    parameters: z.object({ incidentId: z.string() }),
    execute: ({ incidentId }) => {
      toolCalls += 1;
      return `${incidentId} is resolved; owner=platform-observability`;
    },
  });

  const myAgent = agent({
    llm: openai({ model: 'gpt-4o-mini' }),
    tools: [lookupIncident],
  });

  const result = await myAgent.run(
    "Use lookup_incident with incidentId='incident-42', then report the status and owner.",
  );
  if (toolCalls !== 1) throw new Error(`Expected one tool call, received ${toolCalls}`);
  console.log(result.data.result);

  await traceloop.forceFlush(); // ensure spans ship before exit
}

main();

Run it with LOGFIRE_TOKEN=<write-token> OPENAI_API_KEY=<key> npx tsx main.ts, then open the Live view of your project. The example fails unless the native LlamaIndex agent executes lookup_incident; the resulting trace contains the agent run, its model calls, and that tool call.

Managed prompts

You can author and version prompts in Prompt Management and fetch them at runtime with the Logfire TypeScript SDK using defineTemplateVar:

import * as logfire from '@pydantic/logfire-node';
import { defineTemplateVar } from '@pydantic/logfire-node/vars';

logfire.configure({ serviceName: 'llamaindex-ts-logfire' });

const systemPrompt = defineTemplateVar<string, { topic: string }>('prompt__system', {
  default: 'You are a helpful assistant about {{topic}}.',
  templateInputsSchema: {
    type: 'object',
    properties: { topic: { type: 'string' } },
    required: ['topic'],
  },
});

const resolved = await systemPrompt.get({ topic: 'observability' });
// Pass resolved.value into your agent's system prompt.

See Use Prompts in Your Application for the production workflow (promoting versions, rollout targeting).