AI Attributes Audit

Audit how completely your AI telemetry follows the OpenTelemetry GenAI semantic conventions — a native Dynatrace AppEngine app.

AI Attributes Audit scans the gen_ai.* spans your collectors already emit into Dynatrace Grail and, section by section, reports which expected attributes are present versus missing across your fleet — so you can find and close instrumentation gaps before they cost you observability.

Carved out of AI Observability 3.0, where this began life as one of its tabs. It shares that app's global scope model (timeframe + segments + filters + sampling + scan-limit) so audit numbers line up with what you see there.

---

What it does

The OpenTelemetry GenAI conventions define a rich attribute surface for agents, tools, models, prompts, retrieval, evaluation, and platform context — but most fleets emit only a fraction of it. This app is the inventory: it knows the full catalog of attributes worth emitting, probes your spans for each one, and shows you exactly where you stand.

• Coverage overview — an overall coverage ring plus hero stats: attributes present (X/Y), categories complete, categories with gaps, and (sampling-extrapolated) span activity.
• Sectioned catalog — attributes grouped across six areas: Core (LLM/inference gen_ai.*), Orchestration (Agent, Traceloop workflow, LangGraph), Tools (gen_ai.tool.*, MCP mcp.*), Retrieval (vector DB), Quality (evaluation & quality, session & user), and Platform (infrastructure/platform context).
• Live table-of-contents that jumps to and expands any section.
• Per-attribute detail — Present/Missing verdict, a "what it buys you" one-liner, a coverage bar with span count, and a detail modal that links to the canonical OTel / OpenLLMetry specs. A caveat explains that heavy sampling can produce false "missing" readings.

Global scope

Every number on the page honours the toolbar selections, exactly like the parent app:

• Timeframe — header selector that persists per user and travels in the URL.
• Segments — Dynatrace filter-segments (with variables) to slice the fleet by team, service, environment, or deployment.
• Global attribute filter — click-to-filter on any value, with on-demand server-side value discovery for high-cardinality keys.
• Sampling — extrapolate from a sampled span population.
• Scan limit — 500 GB / 1 TB / 2 TB / 5 TB query budget; every query routes through useScopedDql, which rewrites scanLimitGBytes and injects segments + global filters.

Getting started

npm install
npm run start      # dev server
npm run build      # production bundle to dist/
npm run deploy     # deploy to the tenant in app.config.json

Other scripts: npm run typecheck, npm run lint, npm test.

Set your environment in app.config.json (environmentUrl) before deploying.

Required permissions

Declared in app.config.json:

Scope	Why
storage:spans:read	Read spans for attribute-coverage auditing
storage:buckets:read	Grail bucket-level access to the spans table
storage:entities:read	Resolve AppCI entity IDs and service names
storage:lookups:read	Lookup CMDB entities for scope resolution
storage:filter-segments:read	Apply tenant-defined filter segments
storage:files:read	Read /lookups/dynatrace/* for segment variables
storage:smartscape:read	Read smartscape nodes/relations for topology segments
state:user-app-states:read / :write	Persist per-user scan-limit & sampling

Project layout

ui/app/
├── pages/AttributeAudit/   The audit page: catalog, queries, hooks, panels
├── scope/                  Timeframe, segments, scan-limit, sampling, global
│                           filter contexts + the useScopedDql query rewriter
├── layout/                 Global filter strip + sampling/scan-limit controls
├── components/             Header, footer, and shared leaf components
├── data/                   Formatting helpers
├── lib/                    Tenant helpers
└── theme/                  Theme tokens + styles

License

MIT — see LICENSE.