Este guia está disponível atualmente em inglês e chinês.

Agents / JevModel

Route an agent’s next step

Choose between search, calculation, response, and clarification.

Define the available actions

Use Choice labels that correspond to real actions your agent can perform. Include the user request, available tools, and any evidence already present in state. Do not offer “answer” when the agent has no evidence it can cite.

Keep execution in the host

A routing answer is only a recommendation. The host should check permissions, arguments, rate limits, and user approvals before it calls any external tool.

Test a concrete route before enabling tools

The request below asks whether a customer message needs search, clarification or review. Send it through the ordinary JevModel API after email verification; keep the key on your server. The selected label only chooses a permitted branch. Test a request with a known order ID, another with no ID, and one asking for an unsupported action. Log missing outputs and uncertain cases as well as successful labels. Write an explicit host mapping from allowed labels to validated functions; never evaluate model output as code.

Jev probability, confidence and human-review thresholdsAdd a decision gate before tool useAI Decision Flows with Jev and human reviewJevModel API: request and responseAI People Search with Jev: Screen Candidates Against Your ICP Before Paying for Email Lookups
{
  "state": {
    "message": "Where is my order?",
    "order_id": null,
    "available_tools": [
      "read_order_status"
    ]
  },
  "questions": {
    "next_step": {
      "type": "choice",
      "instructions": "Which permitted next step fits the available evidence?",
      "criteria": {
        "search": "Read order status only when a usable order ID is supplied.",
        "clarify": "Ask for the missing order ID before reading status.",
        "review": "Unsupported or ambiguous requests need a human."
      }
    }
  }
}

JevModel é independente e não é afiliado à TypeSafe AI.