LangFuse Trace Analyzer

LangFuse is an open-source LLM observability and analytics platform. It provides tracing for LLM calls, RAG pipelines, and agent workflows — giving you visibility into latency, token usage, cost, and quality scores. It supports the OpenTelemetry standard and integrates with LangChain, OpenAI SDK, Anthropic, and custom code.

A Trace is the top-level container representing a single user request or workflow run. Spans are time-bounded operations within a trace (e.g., a database lookup). Observations are LLM-specific spans: Generation (an LLM call with input/output/tokens) or Event (a discrete non-LLM action like a cache hit). Spans can nest to show call hierarchy.

Option 1: Use the LangFuse SDK directly (langfuse.trace(), langfuse.generation()). Option 2: Use @observe() decorator in Python for automatic tracing. Option 3: Use integration wrappers for LangChain or OpenAI (observe_openai()). Option 4: Use the OpenTelemetry exporter for standard OTEL instrumentation.

Scores let you attach quality metrics to traces. Types: manual (human labels), model-based (LLM-as-judge), and custom (from your app). Common scores: relevance (0-1), faithfulness (0-1), toxicity (bool). Use scores to track quality over time, filter for bad traces, and train feedback loops.

Yes — LangFuse is fully open-source (MIT license). Self-host with Docker Compose or Kubernetes. The cloud version (cloud.langfuse.com) has a free tier. Self-hosted gives you full data control and no per-event pricing. Use environment variables LANGFUSE_SECRET_KEY, LANGFUSE_PUBLIC_KEY, and LANGFUSE_HOST to configure.