Decision Router

The Decision Router node sends a workflow down one of its paths, chosen by a decision model. Each path is an option of a question about a state: the decision model names the option that fits, the node runs the path, and you know how sure the model was.

Node kind: extensions.com.cloud-apim.llm-extension.decision_router

How it works

The router takes a state, a question and its paths: each path has a name and a description of what it stands for
The decision model is asked the question as a choice, with one option per path
It answers with the option that fits best, its confidence and the probability of every option
The path of that option runs, and its result is the result of the router

A decision model always names one of the options it was given, in a fraction of a second and for a fraction of the cost of a text model. Routing no longer depends on a model choosing to call a tool, nor on a sentence to parse.

Configuration

Parameter	Type	Required	Description
`provider`	string	yes	The decision model id
`paths`	array	yes	The paths to choose from, at least two. Each one has an `id`, a `description` and the `node` to run
`state`	string, object or array	no	What the decision is about. The input of the workflow when omitted
`instructions`	string	no	The question asked to the decision model
`model`	string	no	The model to use, the one of the decision model when omitted
`min_confidence`	number	no	Between 0 and 1. Below it, the answer of the decision model is not followed
`default_path`	string	no	The `id` of the path to follow when the decision model is not confident enough
`decision_result`	string	no	The name of the workflow memory that receives the answer of the decision model

The description of a path is what the decision model reads to choose: say what the path is for, the way you would brief someone doing the routing by hand.

Example

{
  "kind": "extensions.com.cloud-apim.llm-extension.decision_router",
  "provider": "decision-model_xxxxx",
  "state": "${input.ticket}",
  "instructions": "Which team should handle this support request?",
  "paths": [
    {
      "id": "billing",
      "description": "Invoices, payments and refunds",
      "node": {
        "kind": "call",
        "function": "extensions.com.cloud-apim.llm-extension.llm_call",
        "args": {
          "provider": "provider_xxxxx",
          "payload": {
            "messages": [
              { "role": "system", "content": "You answer billing questions." },
              { "role": "user", "content": "${input.ticket}" }
            ]
          }
        }
      }
    },
    {
      "id": "technical",
      "description": "Outages, bugs and integrations",
      "node": {
        "kind": "call",
        "function": "extensions.com.cloud-apim.llm-extension.llm_call",
        "args": {
          "provider": "provider_xxxxx",
          "payload": {
            "messages": [
              { "role": "system", "content": "You troubleshoot technical issues." },
              { "role": "user", "content": "${input.ticket}" }
            ]
          }
        }
      }
    }
  ],
  "result": "answer"
}

A path can also be the node itself, carrying its own id and description, the way the paths of the AI Agent Router are written.

Only follow an answer the model is sure of

The decision model tells how confident it is in its choice, between 0 and 1. With min_confidence, an answer below that confidence is not followed: the default_path runs instead, a human review queue for instance. Without a default path, no path runs and the router returns null, like a switch without a match.

{
  "kind": "extensions.com.cloud-apim.llm-extension.decision_router",
  "provider": "decision-model_xxxxx",
  "state": "${input.ticket}",
  "instructions": "Which team should handle this support request?",
  "min_confidence": 0.6,
  "default_path": "human",
  "decision_result": "routing",
  "paths": [
    { "id": "billing", "description": "Invoices, payments and refunds", "node": { "kind": "value", "value": "billing" } },
    { "id": "technical", "description": "Outages, bugs and integrations", "node": { "kind": "value", "value": "technical" } },
    { "id": "human", "description": "Anything else, to be read by a person", "node": { "kind": "value", "value": "human" } }
  ],
  "result": "team"
}

Keep the decision

With decision_result, the answer of the decision model is stored in the workflow memory under that name, for the nodes that follow: to log it, to explain a routing, or to act on the probabilities.

{
  "type": "choice",
  "choice": "technical",
  "confidence": 0.78,
  "probabilities": { "billing": 0.1, "technical": 0.85, "human": 0.05 }
}

Decision Router vs AI Agent Router

	Decision Router	AI Agent Router
Who chooses	A decision model	A text model, by calling a tool
Answer	Always one of the paths, with its confidence and the probability of each path	The tool the model decides to call
Speed and cost	A fraction of a second, a fraction of the cost of a text model	Those of a chat completion
Use it when	The paths are known and you want a fast, measurable choice	The choice needs the reasoning of a text model

The call to the decision model is accounted for like any other: it is audited, priced and counted against the budgets of the decision model.

How it works​

Configuration​

Example​

Only follow an answer the model is sure of​

Keep the decision​

Decision Router vs AI Agent Router​