Langfuse
inari.langfuse
A Docker Compose file for Langfuse alone, generated from the catalog and checked by the safety inspector. No account needed.
Observability platform for language-model applications, recording traces of nested calls with prompts, completions, latency, token counts and cost, plus prompt versioning and evaluation runs.
Langfuse captures a trace per application run with nested spans for retrieval steps, tool calls and model invocations, so a bad answer can be inspected down to the individual call that produced it. Each observation carries its prompt, completion, model, latency and token usage, which the platform prices into cost figures aggregated by user, session or release. Prompt management versions prompts server-side and serves them to applications at runtime, and evaluation runs score traced output against datasets or model-based graders. A self-hosted deployment needs PostgreSQL, ClickHouse, Redis and S3-compatible storage behind it.
You know it worked when
- The interface loads and a project can be created.
- An instrumented application produces a trace that appears in the interface.
- Nested spans within that trace show model, latency and token counts.
- Cost aggregation reports a figure for the traced model.
- A prompt versioned in the platform is retrieved by the application at runtime.
Known sharp edges
- Self-hosting is a multi-service stack, and a failure in ClickHouse or the object store surfaces as missing traces rather than as a clear error.
- Traces contain full prompts and completions, which routinely include the customer data the application was processing.
- Encryption and authentication secrets are generated at first start, and losing them makes existing encrypted fields unreadable.
- Enterprise-gated features appear in the interface and only fail at the point of use, which makes the licensing boundary easy to cross unintentionally.