Write a Recipe

The guide sorts a support inbox. Every page uses the same Recipe, and it grows as you go.

A Recipe is a JSON file. It names the model, sets the review threshold, and lists the Questions. Save this as recipe.json:

{
  "format": 1,
  "name": "support-inbox",
  "version": 1,
  "model": "jev-1.13.0",
  "reviewThreshold": 0.7,
  "questions": {
    "category": {
      "type": "choice",
      "instructions": "Which category best describes this message?",
      "criteria": {
        "complaint": "The sender is unhappy with the product or service",
        "question": "The sender asks how something works",
        "other": "Anything else, including spam and thanks"
      }
    },
    "wants_human": {
      "type": "noul",
      "instructions": "Does the sender ask to speak to a person?"
    }
  }
}

Two Questions so far. category is a Choice: the model picks one of the three labels, and the description beside each label is what it picks on, so write descriptions that separate the labels rather than restate them. wants_human is a yes/no Question, "type": "noul". It has no criteria here, though it can carry a description for true and for false when the boundary needs spelling out.

The fields around the Questions:

• format is the Recipe file format, 1 today. A breaking change to the format bumps it.
• name and version identify the Recipe. Every Decision names both, so change version whenever you change a Question.
• model pins the model that answers. jev-1.13.0 is the current one.
• reviewThreshold is the confidence below which a Decision is Unsure. It is required in the file, so a Recipe on disk never silently gets a default.

Question names are the keys under questions. They become keys in every Decision's answers and in every Fixture's expect, so pick names you are happy to see in JSON and in code.

Next: measure the Recipe against examples you label before it touches real Records.