Normalizing

Slate editors can edit complex, nested data structures. And for the most part this is great. But in certain cases inconsistencies in the data structure can be introduced—most often when allowing a user to paste arbitrary richtext content.

"Normalizing" is how you can ensure that your editor's content is always of a certain shape. It's similar to "validating", except instead of just determining whether the content is valid or invalid, its job is to fix the content to make it valid again.

Built-in Constraints

Slate editors come with a few built-in constraints out of the box. These constraints are there to make working with content much more predictable than standard contenteditable. All of the built-in logic in Slate depends on these constraints, so unfortunately you cannot omit them. They are...

1. All Element nodes must contain at least one Text descendant — even Void Elements. If an element node does not contain any children, an empty text node will be added as its only child. This constraint exists to ensure that the selection's anchor and focus points (which rely on referencing text nodes) can always be placed inside any node. Without this, empty elements (or void elements) wouldn't be selectable.

2. Two adjacent texts with the same custom properties will be merged. If two adjacent text nodes have the same formatting, they're merged into a single text node with a combined text string of the two. This exists to prevent the text nodes from only ever expanding in count in the document, since both adding and removing formatting results in splitting text nodes.

3. Block nodes can only contain other blocks, or inline and text nodes. For example, a paragraph block cannot have another paragraph block element and a link inline element as children at the same time. The type of children allowed is determined by the first child. Any other non-conforming children are tried to be converted (if possible) or removed. This ensures that common richtext behaviors like "splitting a block in two" function consistently. Conversion of block nodes is done by unwrapping the block node; conversion of inline/text nodes is performed by wrapping such nodes into a fallbackElement if specified in the normalizeNode options. The fallbackElement can be specified by editors overriding the normalizeNode function.

4. Inline nodes cannot be the first or last child of a parent block, nor can it be next to another inline node in the children array. If this is the case, an empty text node will be added to correct this to be in compliance with the constraint.

5. The top-level editor node can only contain block nodes. If any of the top-level children are inline or text nodes they will be removed. This ensures that there are always block nodes in the editor so that behaviors like "splitting a block in two" work as expected.

6. Nodes must be JSON-serializable. For example, avoid using undefined in your data model. This ensures that operations are also JSON-serializable, a property which is assumed by collaboration libraries.

7. Property values must not be null. Instead, you should use an optional property, e.g. foo?: string instead of foo: string | null. This limitation is due to null being used in operations to represent the absence of a property.

These default constraints are all mandated because they make working with Slate documents much more predictable.

🤖 Although these constraints are the best we've come up with now, we're always looking for ways to have Slate's built-in constraints be less constraining if possible—as long as it keeps standard behaviors easy to reason about. If you come up with a way to reduce or remove a built-in constraint with a different approach, we're all ears!

Adding Constraints

The built-in constraints are fairly generic. But you can also add your own constraints on top of the built-in ones that are specific to your domain.

To do this, you extend the normalizeNode function on the editor. The normalizeNode function gets called every time an operation is applied that inserts or updates a node (or its descendants), giving you the opportunity to ensure that the changes didn't leave it in an invalid state, and correcting the node if so.

For example here's a plugin that ensures paragraph blocks only have text or inline elements as children:

import { Transforms, Element, Node } from 'slate'

const withParagraphs = editor => {
  const { normalizeNode } = editor

  editor.normalizeNode = (entry, options) => {
    const [node, path] = entry

    // If the element is a paragraph, ensure its children are valid.
    if (Element.isElement(node) && node.type === 'paragraph') {
      for (const [child, childPath] of Node.children(editor, path)) {
        if (Element.isElement(child) && !editor.isInline(child)) {
          Transforms.unwrapNodes(editor, { at: childPath })
          return
        }
      }
    }

    // Fall back to the original `normalizeNode` to enforce other constraints.
    normalizeNode(entry, options)
  }

  return editor
}

This example is fairly simple. Whenever normalizeNode gets called on a paragraph element, it loops through each of its children ensuring that none of them are block elements. And if one is a block element, it gets unwrapped, so that the block is removed and its children take its place. The node is "fixed".

But what if the child has nested blocks?

Multi-pass Normalizing

One thing to understand about normalizeNode constraints is that they are multi-pass.

If you check the example above again, you'll notice the return statement:

if (Element.isElement(child) && !editor.isInline(child)) {
  Transforms.unwrapNodes(editor, { at: childPath })
  return
}

You might at first think this is odd, because with the return there, the original normalizeNodes will never be called, and the built-in constraints won't get a chance to run their own normalizations.

But, there's a slight "trick" to normalizing.

When you do call Transforms.unwrapNodes, you're actually changing the content of the node that is currently being normalized. So even though you're ending the current normalization pass, by making a change to the node you're kicking off a new normalization pass. This results in a sort of recursive normalizing.

This multi-pass characteristic makes it much easier to write normalizations, because you only ever have to worry about fixing a single issue at once, and not fixing every possible issue that could be putting a node in an invalid state.

To see how this works in practice, let's start with this invalid document:

<editor>
  <paragraph a>
    <paragraph b>
      <paragraph c>word</paragraph>
    </paragraph>
  </paragraph>
</editor>

The editor starts by running normalizeNode on <paragraph c>. And it is valid, because it contains only text nodes as children.

But then, it moves up the tree, and runs normalizeNode on <paragraph b>. This paragraph is invalid, since it contains a block element (<paragraph c>). So that child block gets unwrapped, resulting in a new document of:

<editor>
  <paragraph a>
    <paragraph b>word</paragraph>
  </paragraph>
</editor>

And in performing that fix, the top-level <paragraph a> changed. It gets normalized, and it is invalid, so <paragraph b> gets unwrapped, resulting in:

<editor>
  <paragraph a>word</paragraph>
</editor>

And now when normalizeNode runs, no changes are made, so the document is valid!

🤖 For the most part you don't need to think about these internals. You can just know that anytime normalizeNode is called and you spot an invalid state, you can fix that single invalid state and trust that normalizeNode will be called again until the node becomes valid.

Empty Children Early Constraint Execution

One special normalization executes before all other normalizations and this can be important to keep in mind when writing your normalizers.

Before any of the other normalizations can execute, Slate iterates through all Element nodes and makes sure they have at least one child. If it does not, an empty Text descendant is created.

This can trip you up when you have custom handling when an Element has no children. For example, if a table element has no rows, you may wish to remove the table; however, this will never happen because a Text node would automatically be created before that normalization could run.

Incorrect Fixes

One pitfall to avoid is creating an infinite normalization loop. This can happen if you check for a specific invalid structure, but then don't actually fix that structure with the change you make to the node. This results in an infinite loop because the node continues to be flagged as invalid, but it is never fixed properly.

For example, consider a normalization that ensured link elements have a valid url property:

// WARNING: this is an example of incorrect behavior!
const withLinks = editor => {
  const { normalizeNode } = editor

  editor.normalizeNode = (entry, options) => {
    const [node, path] = entry

    if (
      Element.isElement(node) &&
      node.type === 'link' &&
      typeof node.url !== 'string'
    ) {
      // ERROR: null is not a valid value for a url
      Transforms.setNodes(editor, { url: null }, { at: path })
      return
    }

    normalizeNode(entry, options)
  }

  return editor
}

This fix is incorrectly written. It wants to ensure that all link elements have a url property string. But to fix invalid links it sets the url to null, which is still not a string!

In this case you'd either want to unwrap the link, removing it entirely. Or expand your validation to accept an "empty" url == null as well.

Implications for Other Code

Sequences of Transforms may need to be wrapped in Editor.withoutNormalizing if the node tree should not be normalized between Transforms. This is frequently the case when you unwrapNodes followed by wrapNodes. For example, you might write a function to change the type of a block as follows:

const LIST_TYPES = ['numbered-list', 'bulleted-list']

function changeBlockType(editor, type) {
  Editor.withoutNormalizing(editor, () => {
    const isActive = isBlockActive(editor, type)
    const isList = LIST_TYPES.includes(type)

    Transforms.unwrapNodes(editor, {
      match: n =>
        LIST_TYPES.includes(
          !Editor.isEditor(n) && SlateElement.isElement(n) && n.type
        ),
      split: true,
    })
    const newProperties = {
      type: isActive ? 'paragraph' : isList ? 'list-item' : type,
    }
    Transforms.setNodes(editor, newProperties)

    if (!isActive && isList) {
      const block = { type: type, children: [] }
      Transforms.wrapNodes(editor, block)
    }
  })
}