langfuse

Why we included this project

Once an LLM feature ships, the hard part is figuring out why a specific answer came out wrong. Langfuse gives you a self-hosted place to see that: it captures traces from OpenAI, LangChain, LlamaIndex, LiteLLM, and other stacks, and lets you drill into individual calls or whole user sessions. Prompt versions and configs live in one place with client-side caching, so iterating on a prompt doesn't add latency, and the evaluation side covers LLM-as-a-judge, code checks, manual labeling, and user feedback. You can also build datasets for test sets and run structured experiments before deploying, and jump from a bad trace straight into the playground to tweak the prompt. For small teams that want observability and evals without per-seat SaaS pricing, it's a practical option that self-hosts and runs on Kubernetes and the major clouds.