My requests are failing: a quick checklist

Run through these before opening a request. Most issues turn up in the first three steps.

1. Status: check status.neuraltrust.ai for an ongoing incident.

2. Wallet or limits: if you’re on Free or Standard, make sure you haven’t hit your monthly allowance or run out of balance. See Questions about your plan, billing or wallet?

3. The status code:

4. The model provider: the provider itself may be failing, or your provider key may be out of quota. Routingfallbacks keep traffic flowing to another provider.

5. Recent changes: a new policy, a rotated key or a new deployment are often behind sudden errors.

When you open the request, include the request ID, the timestamp with timezone, the gateway, the model and the full error response (with any secrets removed).

Requests blocked with 429 (rate limit or budget reached)

A 429 usually means a limit is doing its job. The response tells you which one.

Plan rate limits

Each plan has a maximum number of requests per minute for TrustGate and TrustGuard. You can see them on the pricing page. If you regularly hit them, consider upgrading.

Rate Limiter policies

These are limits you’ve set yourself: Rate Limiter and Per-Tool Rate Limiter. Retry after the wait time given in the response.

LLM Budget

A token or dollar budget is used up for the current window (LLM Budget). Good to know:

• The window is fixed. It starts with the first charged request and resets all at once. The reset time is in the budget headers.

• The request that crosses the limit is still served. The ones after it are refused.

• Instead of blocking, you can downgrade requests to a cheaper model.

• You can run the policy in Observe mode to tune the limits without blocking anyone.

Tip: to give each end user their own budget, use Group by header (for example X-User-Id) instead of creating one policy per user.