@mayurajs/observability-arize

Arize AX and Phoenix for Mayura's traces (mayura/exporter-otlp), with OpenInference attributes beside the OpenTelemetry GenAI ones.

npm install mayura @mayurajs/observability-arize

import { agentRunTraceSpans } from 'mayura/exporter-otlp';
import { arizeTraceExporter } from '@mayurajs/observability-arize';

const arize = arizeTraceExporter({ spaceId, apiKey, projectName: 'support-agent', region: 'us', serviceName: 'support-agent' });
await arize.sink(agentRunTraceSpans(observer.inspect(handle.id)?.recent ?? []), { signal: AbortSignal.timeout(10_000) });

• Regions us (the default), eu and ca. projectName is a stable identifier, such as support-agent.
• Each span carries openinference.span.kind (AGENT for runs, LLM for model calls, TOOL for tool calls, CHAIN otherwise), the model and provider, token counts, and tool and agent names, so runs show as agent traces. openInferenceAttributes is exported for other OpenInference backends' OTLP exporters (spanAttributes).
• Spans are metadata only, so traces carry no prompts or outputs.

Phoenix

import { phoenixTraceExporter } from '@mayurajs/observability-arize';

// A Phoenix Cloud space; or a self-hosted Phoenix, such as http://127.0.0.1:6006 with allowInsecureLoopback: true.
const phoenix = phoenixTraceExporter({ baseUrl: 'https://app.phoenix.arize.com/s/acme', apiKey, projectName: 'support-agent', serviceName: 'support-agent' });

• Phoenix takes only OTLP's protobuf encoding, so this exporter sends protobuf, with the same OpenInference attributes.
• apiKey is needed when Phoenix's authentication is on (always on Phoenix Cloud).
• The sink also works as the sink of createWorkflowTraceExport, for durable workflow runs.
• Credentials are passed in; nothing is read from the environment. Every other option is the OTLP trace exporter's.

See the observability guide. Apache-2.0.