Integration with MLFlow for Tracing

[meetzuber]

How would you describe the priority of this documentation request

Critical (currently preventing usage)

Describe the future/missing documentation

How to integrate NAT with MLflow for tracing. I did not found any refence in existing document. if MLFlow integration is missing it should be added as feature request.

Thanks

Where have you looked?

https://docs.nvidia.com/nemo/agent-toolkit/latest/run-workflows/observe/observe.html
https://docs.nvidia.com/nemo/agent-toolkit/latest/extend/custom-components/telemetry-exporters.html

Code of Conduct

• I agree to follow this project's Code of Conduct
• I have searched the open documentation issues and have found no duplicates for this bug report

[EnesYilmazcode]

Thanks for filing this. I'd like to pick it up and implement it.

Short version: MLflow 3.6+ exposes a native OTLP/HTTP trace ingestion endpoint (POST /v1/traces), and NAT already ships thin OTLP-target exporters, so this slots into the existing pattern with no custom span translation.

My plan is to mirror the recent Weave OTel exporter (#1827) and follow the same shape as the weave/phoenix exporters: add an mlflow telemetry exporter to the shared opentelemetry plugin that points an OTLP span exporter at the MLflow server's /v1/traces endpoint, passing the experiment id via the x-mlflow-experiment-id header (plus optional auth headers for hosted/Databricks MLflow). Alongside it I'd add an example config under simple_calculator_observability and a docs page under run-workflows/observe, wired into the observe index. It stays a thin wrapper over the existing OTLPSpanAdapterExporter since MLflow speaks standard OTLP/HTTP-protobuf, so there is no MLflow-specific span mapping.

Before I open a PR, does this direction look right, and is there a preference on the config field names (endpoint / experiment id / headers)? I'll sign off my commits per the contributing guide. Happy to adjust scope if you'd rather keep it docs-only.