Track AI Model Executions with LangFuse Observability for Better Performance Insights

About this template

This template is to demonstrate how to trace the observations per execution ID in Langfuse via ingestion API.

Good to know

• Endpoint: https://cloud.langfuse.com/api/public/ingestion
• Auth is a Generic Credential Type with a Basic Auth: username = you_public_key, password = your_secret_key.

How it works

• Trigger: the workflow is executed by another workflow after an AI run finishes (input parameter execution_id).

• Remove duplicates
  Ensures we only process each execution_id once (optional but recommended).

• Wait to get execution data
  Delay (60-80 secs) so totals and per-step metrics are available.

• Get execution
  Fetches workflow metadata and token totals.

• Code: structure execution data
  Normalizes your run into an array of perModelRuns with model, tokens, latency, and text previews.

• Split Out → Loop Over Items
  Iterates each run step.

• Code: prepare JSON for Langfuse
  Builds a batch with:

  • trace-create (stable id trace-<executionId>, grouped into session-<workflowId>)
  • generation-create (model, input/output, usage, timings from latency)

• HTTP Request to Langfuse
  Posts the batch. Optional short Wait between sends.

Requirements

1. Langfuse Cloud project and API keys
2. n8n instance with the HTTP node

Customizing

1. Add span-create and set parentObservationId on the generation to nest under spans.
2. Add scores or feedback later via score-create.
3. Replace sessionId strategy (per workflow, per user, etc.). If some steps don’t produce tokens, compute and set usage yourself before sending.