Model management
Frontend route:- Add or update model providers.
- Enable model entries for agents.
- Sync pricing from the model catalog.
- Disable models that should not be used.
Model create / edit dialog
Add and Edit share one<ModelDialog /> component (mode: "create" | "edit").
- Standard fields: Provider, Model ID (catalog autocomplete), API Key (BYOK), Display Name, Base URL, Description (create only).
- Advanced Settings (collapsed by default): input/output cost per 1M tokens, context window, max output tokens, and capability switches (Vision, Function Calling, Streaming, Active).
- Provider and Model ID stay editable in both modes.
Test a model connection
Use Test Connection before saving to verify that the selected provider, model ID, API key, and base URL can complete a lightweight model request.1
Configure the model
Select a provider and model ID, then select an API key and enter a base URL
when the provider requires one.
2
Run the connection test
Select Test Connection. The button is available after you enter a model
ID.
3
Review the result
Continue saving after a successful test. If the test fails, correct the
credentials, model ID, provider, or base URL shown in the error and test
again.
A successful connection test confirms that the current configuration can reach
the provider. It does not save or enable the model.
SystemOne API Gateway Reverse-Proxy
Directly evaluate input state against bounded questions using the TypeSafe Jev SystemOne reasoning engine via Pinter API Gateway:Security Labels and Intelligence Ratings (L1 Only)
Pinter models carry enterprise security classifications and intelligence ratings curated by Platform Architects (L1) viaPATCH /api/agents/models/{model_id}/curation.
Security Labels
Security labels determine where and how a model may be utilized across departments:Intelligence Ratings (1 to 5 Bulbs)
Displayed in model pickers as bulb indicators (1 to 5 bulbs) to guide creators toward cost-effective models:- 1–2 Bulbs (Lightweight / Fast): Small distilled models (e.g., Flash, Haiku, Mini) ideal for routing, extraction, and simple classification.
- 3 Bulbs (Balanced): Standard workhorse models for multi-step reasoning.
- 4–5 Bulbs (Advanced Reasoning): Flagship frontier models (e.g., Claude 3.5 Sonnet, GPT-4o, o1, Gemini 1.5 Pro) with chain-of-thought capabilities.
Provider API key management
Provider credentials are managed from the Models page. Each API key is a parent record: one key can serve many models, and models are assigned to a key from the model side. Admin key endpoints:- Credentials are encrypted at rest and shown masked after creation.
- Deleting a key does not delete its models; they fall back to the platform default credential.
- Keys marked as BYOK (bring your own key) belong to a specific application or user context.
These provider keys are different from personal platform API keys, which authenticate users calling the Pinter API. See Platform API keys.
Quota management
Frontend route:Default quota on user creation
When a new user is created (via self-registration, SSO JIT provisioning, or admin creation), a default monthly quota is automatically assigned:
Additional approved top-up requests increase the monthly budget.
6-hour sliding budget rate limiter
To prevent rapid budget consumption, each AI request is checked against a sliding 6-hour window:- Budget consumed in the last 6 hours is computed from
token_usage. - Limit = 10% of total monthly budget (
max_cost_per_month + approved_amount). - If the 6-hour spend reaches the limit, the request is blocked with HTTP 429 and message:
"Batas penggunaan 6 jam terlampaui (Maksimal 10% dari total kuota bulanan)."
- 6h Budget (10%) progress bar.
- Near Limit orange badge when >80% consumed.
- Remaining budget until the next sliding window.
Anomaly detection
After each AI request is logged, a background task compares the user’s spend in the last hour with their own usage pattern:- The baseline is the median spend of the user’s active hours over the previous 7 days, excluding the hour being tested. Idle hours and one-off outlier days do not distort it.
- An alert is raised only when the last hour is above 3× the baseline, above the user’s own P95 hour, and at least a minimum amount (default
$0.25). Users with fewer than 6 active hours of history are skipped. - The multiple of baseline is stored in
quota_alerts.threshold_percent(for example840means 8.4× normal) and sets the severity: Elevated below 5×, High from 5×, Critical from 10×. - If an unacknowledged alert already exists for the quota, it is escalated in place when the severity increases; otherwise it is left untouched.
ANOMALY_RATIO_THRESHOLD, ANOMALY_MIN_COST_USD and ANOMALY_MIN_ACTIVE_HOURS environment variables.
Admins review alerts in the Anomaly tab of the Usage page. It shows summary counts, an Open/All filter (history of acknowledged alerts), and an expandable detail per alert: last-hour cost versus normal and threshold, whether the spike came from request volume or request size, a 48-hour spend chart, and the top agents, models, API keys and sources in the spike hour. Details are computed from GET /api/governance/quotas/alerts/{alert_id}/anomaly-details. Users with quota:manage can Acknowledge an alert.
Cost tracking — user × model breakdown
The/api/governance/usage/detailed endpoint now returns an additional by_user_model array, showing cost and request count grouped by each user-model combination. This is displayed in the Admin Usage page under the User×Model tab.
Usage/me — 6-hour budget fields
The/api/governance/usage/me response includes additional fields: