"Evaluate with your existing prod data" currently means Langfuse (v4 observations-first export, verified lossless) or hand-shaped JSON. Widening ingest widens who can use the tool.
What's needed, roughly in demand order
- OTel GenAI semantic-convention ingest (OTLP file/collector export) — covers many tracing stacks at once.
- OpenAI request logs / stored completions export.
- Braintrust, LangSmith, Phoenix/Arize exporters as demand appears.
- Each importer proves itself the same way Langfuse did: a real export imports losslessly (contract graduation rule).
The trace contract is the stable target; each importer is a normalizer onto it.
"Evaluate with your existing prod data" currently means Langfuse (v4 observations-first export, verified lossless) or hand-shaped JSON. Widening ingest widens who can use the tool.
What's needed, roughly in demand order
The trace contract is the stable target; each importer is a normalizer onto it.