Token Budgets¶
What tokens are¶
When you send a question to an AI provider, the text is broken into small chunks called tokens (roughly one token per 3–4 characters of text). The AI reads those tokens and generates a response made of more tokens. Cloud AI providers (Anthropic, OpenAI) charge by the number of tokens processed.
For local providers (Ollama, llama.cpp), there's no per-token cost — but setting a token budget still limits computation time.
Setting a token limit¶
In Settings → AI → Token Budget, set a monthly token limit. Once the limit is reached, AI requests stop (by default) until the next month.
Token limits are per household, not per user.
Setting a cost limit¶
Set a monthly cost limit in your home currency. WDIAG tracks the estimated cost of each AI request (provider rate × tokens used) and stops requests when the limit is reached.
Cost tracking is an estimate — actual billing from your provider may differ slightly from WDIAG's estimate.
Overage behavior¶
What happens when a limit is reached:
| Mode | Behavior |
|---|---|
| Block (default) | AI requests are refused. Users see "AI budget reached for this month." |
| Warn and continue | A warning is logged, the request proceeds anyway |
| Silent | The request proceeds without any warning |
Change overage behavior in Settings → AI → Token Budget.
Monitoring usage¶
The AI section of your settings shows usage for the current month: - Tokens used / limit - Estimated cost / limit (if a cost limit is set) - Breakdown by provider, model, and insight category (anomaly detection, Q&A, categorization assistance, etc.)
Why budgets exist¶
Cloud AI providers charge real money. A household running active AI insights and asking frequent questions can accumulate noticeable API costs over a month. The default block behavior ensures you don't get surprised at the end of the month.
For local providers, there's no direct cost, but you can still set a token budget to limit how much AI computation runs (relevant on low-power servers where AI inference competes with other resources).
If you're using a local model only and have no cost concerns, you can set no limit at all — just don't configure a budget.