VoltAgent/voltagent @voltagent/langfuse-exporter@1.1.0

on GitHub

Minor Changes

• #554 3a70b05 Thanks @omeraplak! - feat: add createLangfuseSpanProcessor helper and robust attribute mappings for new OpenTelemetry observability

  What changed for you

  • New helper: createLangfuseSpanProcessor to plug Langfuse export directly into VoltAgent’s OpenTelemetry-based observability without touching core.
  • Improved attribute mappings with careful fallbacks to align @voltagent/core span attributes and Langfuse fields (usage, model params, input/output, user/session, tags, names).
  • Updated examples/with-langfuse to demonstrate the new integration.

  Quick start

  import { Agent, VoltAgent, VoltAgentObservability } from "@voltagent/core";
  import { createLangfuseSpanProcessor } from "@voltagent/langfuse-exporter";
  
  // Configure Observability: add Langfuse via SpanProcessor
  const observability = new VoltAgentObservability({
    spanProcessors: [
      createLangfuseSpanProcessor({
        publicKey: process.env.LANGFUSE_PUBLIC_KEY,
        secretKey: process.env.LANGFUSE_SECRET_KEY,
        baseUrl: process.env.LANGFUSE_BASE_URL, // e.g. https://cloud.langfuse.com or self-hosted
        debug: true, // optional
        // batch: { maxQueueSize, maxExportBatchSize, scheduledDelayMillis, exportTimeoutMillis }
      }),
    ],
  });
  
  const agent = new Agent({
    name: "Base Agent",
    // ...model, tools, memory
  });
  
  new VoltAgent({
    agents: { agent },
    observability,
  });

  Environment variables

  • LANGFUSE_PUBLIC_KEY
  • LANGFUSE_SECRET_KEY
  • LANGFUSE_BASE_URL (optional; defaults to Langfuse cloud if omitted)

  Mapping details (highlights)

  • Usage tokens: gen_ai.usage.* ← fallbacks to usage.prompt_tokens, usage.completion_tokens, usage.total_tokens from core.
  • Model params: prefers gen_ai.request.*, falls back to ai.model.* from core.
  • Input/output (generation): prefers ai.prompt.messages / ai.response.text, falls back to generic input / output set by core.
  • Input/output (tools): prefers tool.arguments / tool.result, falls back to input / output.
  • User/session: enduser.id ← user.id, session.id ← conversation.id.
  • Tags: reads tags or parses JSON from prompt.tags if present.
  • Name: prefers voltagent.agent.name then entity.name then span name.

  Example updated

  • See examples/with-langfuse for a complete, working setup using createLangfuseSpanProcessor.