Honest recommendations plus opt-in request enforcement. Target verification records rule matches without changing model traffic. Active exact or semantic cache, deny, retry-cap, evidence-gated model or route, and fallback rules run only for calls wrapped by @traice/sdk.
New rules start in draft. Target verification records which requests match without sending traffic to the candidate. Active supported actions can change calls explicitly wrapped by @traice/sdk.
Showing the 5 newest of 8 rules.
| Name | When | Then | State | Actions |
|---|---|---|---|---|
| Fallback when the primary provider is unavailable | feature = ticket-summary | fallback → gpt-4o | active · fallback | |
| Block requests beyond tenant policy | feature = chat-resolver | deny | active · deny | |
| Cap repeated knowledge retrieval | feature = knowledge-agent | cap retries → gpt-5.5 | active · cap retries | |
| Reuse semantically equivalent knowledge answers | feature = knowledge-agent | cache semantic | active · cache semantic | |
| Reuse exact helpdesk responses | feature = chat-resolver | cache exact | active · cache exact |
| Rule | State | Decisions | Cache hit rate | Projected savings | Realized savings |
|---|---|---|---|---|---|
| Fallback when the primary provider is unavailable | active | 75 | - | $0 | $0 |
| Block requests beyond tenant policy | active | 120 | - | $0 | $720 |
| Cap repeated knowledge retrieval | active | 240 | - | $0 | $960 |
| Reuse semantically equivalent knowledge answers | active | 180 | 100% | $0 | $1,080 |
| Reuse exact helpdesk responses | active | 600 | 80% | $0 | $2,400 |