Doing JSON structured output with Mini Me

So, what is structured output? It's the ability of a model to produce an output/answer in JSON while following a structure defined beforehand, usually JSON described by a JSON Schema. Of course, your model has to support this feature. You can read, for example, OpenAI's documentation on the subject: Structured model outputs.

Defining the structure

With Mini Me, you define this structure using the same notation as Golang structs (yes, I love Go, and it's easy to read), like this for example:

yaml

provider: llmman
model: hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_M
baseUrl: http://127.0.0.1:17434/v1
# From a Docker sandbox, the daemon runs on the host.
fallback: http://host.docker.internal:17434/v1

bashTool: false
editTools: false
planTool: false
fetchContentTool: false
askUserTool: false

acp:
  name: pizza
  title: 🍕 Pizza (JSON)
  
system: |
  You are a pizzaiolo. 
  For the pizza the user names, list its ingredients for one 30 cm pizza.

structuredOutput: |
  // One ingredient of the pizza.
  type Ingredient struct {
      Name       string   `json:"name"`       // the ingredient's name, in the user's language
      Quantity   float64  `json:"quantity"`   // amount for one pizza
      Unit       string   `json:"unit"`       // "g", "ml", "piece", "pinch" or "to taste"
      Category   string   `json:"category"`   // "base", "sauce", "cheese", "meat", "vegetable", "herb" or "other"
      Optional   bool     `json:"optional"`
  }

  type Pizza struct {
      Name        string       `json:"name"`
      Vegetarian  bool         `json:"vegetarian"`
      Ingredients []Ingredient `json:"ingredients"`
      BakingTime  *int         `json:"bakingTime"` // minutes, null when unknown
  }

  var Recipe Pizza

sampling:
  temperature: 0.0
  chat_template_kwargs:
    enable_thinking: false

So here, I want to get the ingredients needed to make a pizza, in JSON format.

Getting the ingredients of a pizza

I haven't mentioned it yet, but you can run Mini Me in "one-shot" or "headless" mode, meaning that mm runs a single prompt and then stops: mm --config agent.yaml --p "tell me something about the hawaiian pizza" or mm --p "tell me something about the hawaiian pizza" agent.yaml

So for our example, we'll use the following command:

bash

mm -p "hawaiian pizza" agent.pizza.yaml | jq .

For a result like this:

json

{
  "bakingTime": 12,
  "ingredients": [
    {
      "category": "base",
      "name": "Pizza Dough",
      "optional": false,
      "quantity": 250,
      "unit": "g"
    },
    {
      "category": "sauce",
      "name": "Tomato Sauce",
      "optional": false,
      "quantity": 60,
      "unit": "ml"
    },
    {
      "category": "cheese",
      "name": "Mozzarella Cheese",
      "optional": false,
      "quantity": 150,
      "unit": "g"
    },
    {
      "category": "meat",
      "name": "Ham",
      "optional": false,
      "quantity": 100,
      "unit": "g"
    },
    {
      "category": "vegetable",
      "name": "Pineapple Chunks",
      "optional": false,
      "quantity": 50,
      "unit": "g"
    },
    {
      "category": "herb",
      "name": "Oregano",
      "optional": false,
      "quantity": 1,
      "unit": "pinch"
    }
  ],
  "name": "Hawaiian Pizza",
  "vegetarian": false
}

How to use the "pizzaiolo" agent in TUI mode?

You can ask the same thing with the TUI, but an agent in "structured output" mode can't use tools, so you won't be able to, for example, save the generated information. The trick to work around this is to use the builtin ask_agent tool. So create a new agent.yaml configuration file, which will be a coordinator agent:

yaml

provider: llmman
model: hf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_M
baseUrl: http://127.0.0.1:17434/v1
fallback: http://host.docker.internal:17434/v1

# The agents Bob can ask with the `ask_agent` tool. 
# Its answer — the JSON — comes back to him as the tool's result.
mmConfigs:
  pizza:
    path: agent.pizza.yaml
    description: All about pizzas.

# Bob reads and writes files (read_file, write_file, edit_file); no shell.
bashTool: false
editTools: true
# planTool: false
# skillsDir: no-skills

yolo: true

acp:
  name: bob
  title: 🤓 Bob (coordinator)

system: |
  Your name is Bob. You coordinate one agent that answers in JSON, through your `ask_agent` tool:
  - `pizza` for pizza recipes;
  When a question is about pizza, call `pizza` agent,
  Call the agent, then work from its JSON.

  To read a file, use `read_file`; to create one, `write_file`.

sampling:
  temperature: 0.0
  parallel_tool_calls: false
  max_tokens: 8192
  chat_template_kwargs:
    enable_thinking: true
    reasoning_effort: low
    
maxTurns: 20
watchdogTimeout: 2m
firstTokenTimeout: 3m