lmolkova

Designing Voice Agents in OpenTelemetry

How to trace speech‑to‑speech (voice‑native) agents like Gemini Live and the OpenAI Realtime API, and why the common modeling choices differ.

Prior art: Arize / OpenInference · LangSmith voice fundamentals · OTel GenAI semconv PR #390

lmolkova

---

name: migrate-to-definitions-schema-v2 description: 'Use when migrating v1 schema files to v2 (file_format: definition/2) for entities, spans, metrics, events, attribute groups, or signal refinements in any semantic-conventions registry. Enforces minimal attribute groups, internal-by-default visibility, flat structure, and refinement of generic signals over re-declaration.

lmolkova

attributes:
- key: android.app.state
  type:
    members:
    - id: created
      value: created
      brief: |
        Any time before Activity.onResume() or, if the app has no Activity, Context.startService() has been called in the app for the first time.
      stability: development
    - id: background

lmolkova

attributes:
- key: android.app.state
  type:
    members:
    - id: created
      value: created
      brief: |
        Any time before Activity.onResume() or, if the app has no Activity, Context.startService() has been called in the app for the first time.
      stability: development
    - id: background

lmolkova

Here're some items that we've considered in the past to be useful in azure-core or clientcore related to observability, but never got to it.

• [Azure Core only] Make slf4j dependency optional - backport Slf4JLoggerShim. Related to Azure/azure-sdk-for-java#38421

• [Azure Core only] Merge tracing and metrics plugins into one - Azure/azure-sdk-for-java#41436 or backport plugin-free OTel support from clientcore

• Allow adding arbitrary key-value-pairs on instrumentation context and stamping them on all nested logs and spans (similar to MDC):

lmolkova

TODOs:

• AI weaver - weaver generate should generate new schema behind feature flag - Liudmila will create an issue

Phases:

1. ingest new format
2. resolve new format - need to advertise and socialize
3. modify resolution to be on the new format

| signal/object | identity property in semsonv | in proto |

lmolkova

package live_check_advice

import rego.v1

required_attributes := [
    "http.request.method",
    "server.address",
    "server.port",
    "url.full",
]

lmolkova

from enum import Enum
import json
from typing import Annotated, Any, List, Literal, Optional, Union
from pydantic import BaseModel, Field, RootModel

class TextPart(BaseModel):
    """
    Describes text content sent to or received from the model.
    """
    type: Literal['text']

lmolkova

RequestOptions scenarios

These are the scenarios we support today, not necessarily the ones we have to provide in the final design.

1. Users can provide static (for a group of operations) configuration. E.g.: configure custom logger for all requests peformed in scope of this API call

2. Provide dynamic (per-request) configuration. E.g. set request body or operation-specific context

So RequestOptions is a mix of (usually) per-request and (usually) static config.

lmolkova

Observability in clientcore

• Key changes
• Champion scenarios: user app
  • Default experience
  • Default experience: I want to use custom correlation-id
  • Default experience: add HTTP logs
  • Experience with OTel + logback + HTTP logs
• Champion scenarios: client lib
• Next steps