Langfuse

Open-source observability and prompt management platform for LLM applications.

Prompt Management

Versioning

Every prompt is versioned with an incrementing number. Prompts can be labeled (e.g., "production", "staging") for environment-specific deployment.

Prompt Composition

Prompts support variable interpolation with double curly braces. Compose complex prompts from reusable segments.

Observability

Traces

Every LLM operation creates a trace — a hierarchical view of the full execution from input to output. Traces contain generations (individual LLM calls) and spans (logical groupings).

Cost Tracking

Token usage and estimated cost are tracked per generation. Aggregate by trace, user, or time period to monitor spend.

Evaluation

Score outputs manually or programmatically. Use scores to identify quality regressions and compare prompt versions.

Best Practices

• Never hardcode prompts in application code — manage them in Langfuse
• Always include token usage in generation spans for accurate cost tracking
• Use trace metadata to link LLM operations back to business context
• Flush traces before task completion to avoid data loss