Generic OTLP/OpenTelemetry provider for AI Assistant telemetry

General
,
1

Hi Directus team,

I’m looking into AI Assistant observability for a self-hosted Directus setup and wanted to ask whether you’d be open to supporting a generic OTLP/OpenTelemetry telemetry provider for the AI Assistant.

From the current docs and source, AI telemetry supports:

AI_TELEMETRY_ENABLED=true
AI_TELEMETRY_PROVIDER=langfuse # or braintrust
AI_TELEMETRY_RECORD_IO=true

That works well for Langfuse/Braintrust, but we would like to send AI Assistant traces to Arize Phoenix. Phoenix can ingest OTLP traces and is commonly used with OpenInference/Vercel AI SDK spans.

A generic provider might look something like:

AI_TELEMETRY_ENABLED=true
AI_TELEMETRY_PROVIDER=otlp
AI_TELEMETRY_OTLP_ENDPOINT=http://phoenix:6006/v1/traces
AI_TELEMETRY_RECORD_IO=true

Optionally, it could also support headers/auth for hosted collectors:

AI_TELEMETRY_OTLP_HEADERS='{"Authorization":"Bearer ..."}'

Use case:

  • capture AI Assistant traces
  • inspect tool calls, especially items queries
  • debug whether the assistant is grounding answers in CMS help articles
  • track latency/token usage/model metadata
  • optionally record full prompt/response IO in controlled environments

Would the Directus team be open to a PR adding either:

  1. a generic OTLP/OpenTelemetry AI telemetry provider, or
  2. a Phoenix-specific provider using Arize/OpenInference packages?

My instinct is that a generic OTLP provider would be more broadly useful than a Phoenix-specific integration, but I wanted to ask before spending time on an implementation.

Thanks!