Langfuse
title: Langfuse type: platform created: 2026-05-26 last_updated: 2026-05-26 related: ["radar/techniques/LLMEvaluationMethodology", "radar/techniques/YearBuildingWithLLMs", "radar/platforms/R2R", "radar/tools/Ollama"] sources: ["https://github.com/langfuse/langfuse"] radar_quadrant: Platforms radar_ring: Assess radar_position: inner
Open-source LLM engineering platform providing observability, metrics, evaluation, prompt management, and datasets for LLM applications. Self-hostable via Docker Compose; integrates with LlamaIndex, LangChain, OpenAI SDK, LiteLLM, and more.
LLM application pipelines are opaque by default. A single user request may involve retrieval, prompt construction, one or more LLM calls, and post-processing. When an answer is wrong or latency is high, there is no native visibility into which step failed, what the prompt contained, or how long each step took. Langfuse brings the tracing and observability model from application performance monitoring to LLM pipelines.
Observability — Every request becomes a trace with nested spans. Each span captures the input, output, latency, token count, and cost for one step (retrieval, prompt render, LLM call, etc.). Traces are queryable and filterable in the Langfuse UI.
Evaluation — Scores can be attached to traces manually, via LLM-as-judge, or via user feedback signals. Scores accumulate into metrics tracked over time, enabling regression detection across model or prompt changes.
Prompt management — Prompts are stored centrally in Langfuse, version-controlled, and fetched by the application at runtime. Changes to prompts are tracked and their effect visible in traces without a code deployment.
Datasets — Curated input/output pairs collected from traces form evaluation datasets. A dataset run applies a new prompt or model version against the dataset and scores the results, enabling systematic before/after comparison.
Langfuse integrates via:
- Python and TypeScript SDKs with decorators for automatic span creation
- OpenTelemetry-compatible tracing for language-agnostic instrumentation
- Environment variable injection for zero-code integration with LiteLLM and some frameworks
Self-hosted deployment runs via Docker Compose with a PostgreSQL backend. A managed cloud option is available.
Langfuse sits in the Assess ring of the Platforms quadrant, at inner position. First studied via the GitHub repository (2024-08-03). Fills the LLM observability slot absent from the radar — a production requirement for any LLM application beyond prototype stage. Inner position reflects Docker Compose self-hosted deployment, SDK integration with tools already on the radar (LlamaIndex, LiteLLM, OpenAI SDK), and direct complementarity with radar/techniques/LLMEvaluationMethodology and radar/techniques/YearBuildingWithLLMs already assessed. Remaining gate before Trial is a deployed Langfuse instance with traces flowing from at least one production LLM pipeline and one evaluation score type configured.
