Observability (Langfuse)

The workflow and index job can integrate with Langfuse for tracing and prompt management.

Features

LLM Tracing

All calls to language models are traced using Langfuse's LangChain integration. This includes:

• Input and output text.
• Token counts and costs.
• Latency and execution steps.
• Metadata such as session IDs for grouping related traces.

Prompt Management

Prompt templates can be managed directly in the Langfuse UI. The workflow loads prompts from app/observe/langfuse.py by name and label, which allows prompt updates without redeploying the service.

Integration Details

LangfuseClient

This client handles:

• callback initialization for LangChain
• prompt fetching by name and label
• session tracking via request or Kafka session IDs
• dedicated feedback traces for StaticAnswer and NoAnswerPossible responses so internal notes can always point to a Langfuse trace

Environment Variables

Langfuse is typically configured using standard environment variables:

• LANGFUSE_PUBLIC_KEY
• LANGFUSE_SECRET_KEY
• LANGFUSE_HOST

Implementation in Triage

The triage and answer flows wrap chain execution in traces when langfuse_enabled is true.

For internal feedback notes, the workflow also creates explicit Langfuse traces for static answers and no-answer results. That keeps feedback links available even when no LLM-generated trace exists.

These feedback traces reuse the session ID from the related triage run so categorization and feedback stay grouped together in Langfuse.