langfuse-observability

Langfuse Observability

Overview

Langfuse is the open-source LLM observability platform recommended for tracing, monitoring, evaluation, and prompt management. Unlike LangSmith (deprecated), Langfuse is self-hosted, free, and designed for production LLM applications.

---

Core Features

1. Distributed Tracing

Track LLM calls across your application with automatic parent-child span relationships.

from langfuse.decorators import observe, langfuse_context

@observe()  # Automatic tracing
async def analyze_content(content: str, agent_type: str):
    """Analyze content with automatic Langfuse tracing."""

    # Nested span for retrieval
    @observe(name="retrieval")
    async def retrieve_context():
        chunks = await vector_db.search(content)
        langfuse_context.update_current_observation(
            metadata={"chunks_retrieved": len(chunks)}
        )
        return chunks

    # Nested span for generation
    @observe(name="generation")
    async def generate_analysis(context):
        response = await llm.generate(
            prompt=f"Context: {context}\n\nAnalyze: {content}"
        )
        langfuse_context.update_current_observation(
            input=content[:500],
            output=response[:500],
            model="claude-sonnet-4-20250514",
            usage={
                "input_tokens": response.usage.input_tokens,
                "output_tokens": response.usage.output_tokens
            }
        )
        return response

    context = await retrieve_context()
    return await generate_analysis(context)

Result in Langfuse UI:

analyze_content (2.3s, $0.045)
├── retrieval (0.1s)
│   └── metadata: {chunks_retrieved: 5}
└── generation (2.2s, $0.045)
    └── model: claude-sonnet-4-20250514
    └── tokens: 1500 input, 1000 output

2. Token & Cost Tracking

Automatic cost calculation based on model pricing, with a built-in pricing database for all major models and custom pricing support. See references/cost-tracking.md for detailed patterns, dashboard queries, and best practices.