Design a simple JSON schema for structure extraction via LLMs

[vitawasalreadytaken]

We will constrain the LLM output to this schema so we can parse the output reasonably.

First draft, not discussed with anyone yet so there are probably issues with it:

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "DocNode",
  "type": "object",
  "properties": {
    "label": {
      "type": "string",
      "description": "Label of the node, e.g. '1.2' or '1bis' or 'a)' etc. May be empty."
    },
    "type": {
      "type": "string",
      "enum": ["document", "heading", "content", "list", "list_item", "footnote", "image"],
      "description": "Type of the node."
    },
    "content": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "description": "Array of content strings. Usually only one string, but may be more in case of multiple paragraphs."
    },
    "children": {
      "type": "array",
      "items": {
        "$ref": "#"
      },
      "description": "Array of nested nodes."
    }
  },
  "required": ["label", "type"],
  "additionalProperties": false
}

And an example document for that:

{
  "label": "",
  "type": "document",
  "content": [],
  "children": [
    {
      "label": "1",
      "type": "heading",
      "content": ["Introduction"],
      "children": [
        {
          "label": "",
          "type": "content",
          "content": ["This is the first paragraph under the introduction."],
          "children": []
        },
        {
          "label": "",
          "type": "list",
          "content": [],
          "children": [
            {
              "label": "a)",
              "type": "list_item",
              "content": ["This is a point in a list under the first paragraph."],
              "children": []
            },
            {
              "label": "b)",
              "type": "list_item",
              "content": ["This is another point in the list."],
              "children": []
            }
          ]
        },
        {
          "label": "1.1",
          "type": "heading",
          "content": ["Subsection"],
          "children": [
            {
              "label": "",
              "type": "content",
              "content": ["Some subsection content."],
              "children": []
            },
          ]
        }
      ]
    }
  ]
}