Tagging requests for per-feature cost attribution

The single most important habit for getting value out of Flaxxa AI's cost ledger is tagging every request. The tags are how spend gets attributed.

What to tag

At minimum, every request should carry:

• feature — the product feature or use case the call belongs to (e.g. support-bot, doc-summary, code-completion).
• tenant — if you're running multi-tenant SaaS, the tenant id of the customer the call serves.
• user — the end user id, if relevant. Useful for per-user rate limiting and cost charge-back.

Anything else useful — region, language, plan_tier, experiment_id. The schema is open; the dashboard pivots on whatever you set.

How to send tags

Two equivalent options:

Option 1: header (recommended)

X-Flaxxa-Tags: feature=support-bot,tenant=tenant_42,user=u_1842

Option 2: request body

{
  "model": "auto",
  "messages": [...],
  "tags": {
    "feature": "support-bot",
    "tenant": "tenant_42",
    "user": "u_1842"
  }
}

Both land in the ledger identically. The header form is slightly easier to add via a SDK middleware — the SDK can derive tags from the request context once and stop product squads from forgetting.

What appears in the dashboard

The cost ledger pivots by any tag. From AI → Ledger you can:

• Group by feature — see per-feature spend over time
• Group by tenant — for white-label SaaS, see per-customer spend (and charge back if pricing is metered)
• Group by user — find the heaviest users
• Cross-pivot — feature × tenant tells you which features are big drivers per customer

Budget alerts on tags

Set budgets at any tag granularity. AI → Budgets → New. Common patterns:

• Total workspace daily budget (kill switch if anything goes wrong)
• Per-feature monthly budget (so a runaway feature can't drain the workspace)
• Per-tenant daily budget (for multi-tenant SaaS — isolates noisy tenants)

Alerts at 80% of cap (Slack/email); 429 responses at 100%. Soft caps recover at midnight in the budget's timezone.

Common tag-discipline mistakes

• Forgetting to tag legacy code paths. Every untagged call lands as (untagged) on the dashboard. If that bucket is more than 10%, you have squads not tagging.
• Tagging at the wrong granularity. If feature=app for everything, you can't pivot. Use feature=support-bot, not feature=app.
• Mixing case + spelling. support-bot and SupportBot are different tags. Pick a convention; lint for it in CI.