Contribute to this guideReport an issue

guideEditing engine

The @ckeditor/ckeditor5-engine package is by far the biggest package of all. Therefore, this guide will only scratch the surface here by introducing the main architecture layers and concepts. More detailed guides will follow.

# Overview

The editing engine implements an MVC architecture. The shape of it is not enforced by the engine itself but in most implementations it can be described by this diagram:

Diagram of the engine’s MVC architecture.

What you can see are three layers: model, controller and view. There is one model document which is converted to two views — the editing view and the data view. These two views represent, respectively, the content that the user is editing (the DOM structure that you see in the browser) and the editor input and output data (in a format which the plugged data processor understands). Both views feature virtual DOM structures (custom, DOM-like structures) on which converters and features work and which are then rendered to the DOM.

The green blocks are the code introduced by editor features (plugins). So features control what changes are done to the model, how they are converted to the view and how the model needs to be changed based on fired events (view’s and model’s ones).

Let’s now talk about each layer separately.

# Model

The model is implemented by a DOM-like tree structure of elements and text nodes. Unlike in the DOM, in the model, both, elements and text nodes can have attributes.

Like in the DOM, the model’s structure is contained within a document which contains root elements (the model, as well as the view, may have multiple roots). The document also holds its selection and the history of its changes.

Finally, the document, its schema and document markers are properties of the Model. Instance of the Model class is available in the editor.model property. The model, besides holding the properties described above, provides also the API for changing the document and its markers.

editor.model;                       // -> The data model
editor.model.document;              // -> The document
editor.model.document.getRoot();    // -> Document's root
editor.model.document.selection;    // -> Document's selection
editor.model.schema;                // -> Model's schema

# Changing the model

All changes in the document structure, of the document’s selection and even creation of elements, can only be done by using the model writer. Its instance is available in change() and enqueueChange() blocks.

// Inserts text "foo" at the selection position.
editor.model.change( writer => {
    writer.insertText( 'foo', editor.model.document.selection.getFirstPosition() );
} );

// Apply bold to the entire selection.
editor.model.change( writer => {
    for ( const range of editor.model.document.selection.getRanges() ) {
        writer.setAttribute( 'bold', true, range );
} );

All changes done within a single change() block are combined into one undo step (they are added to a single batch). When nesting change() blocks, all changes are added to the outermost change() block’s batch. For example, the below code will create a single undo step:

editor.model.change( writer => {
    writer.insertText( 'foo', paragraph, 'end' ); // foo.

    editor.model.change( writer => {
        writer.insertText( 'bar', paragraph, 'end' ); // foobar.
    } );

    writer.insertText( 'bom', paragraph, 'end' ); // foobarbom.
} );

All changes made to the document structure are done by applying operations. The concept of operations comes from Operational Transformation (in short: OT), a technology enabling collaboration functionality. Since OT requires that a system is able to transform every operation by every other one (to figure out the result of concurrently applied operations), the set of operations needs to be small. CKEditor 5 features a non-linear model (normally, OT implementations use flat, array-like models while CKEditor 5 uses a tree structure), hence the set of potential semantic changes is more complex. To handle that, the editing engine implements a small set of operations and a bigger set of “deltas” — groups of operations with additional semantics attached. Finally, deltas are grouped in batches. A batch can be understood as a single undo step.

# Text attributes

Text styles such as “bold” and “italic” are not kept in the model as elements but as text attributes (think — like element attributes). The following DOM structure:

    "Foo "

would translate to the following model structure:

    "Foo "  // text node
    "bar"   // text node with bold=true attribute

Such representation of inline text styling allows to significantly reduce the complexity of algorithms operating on the model. For instance, if you have the following DOM structure:

    "Foo "

and you have a selection before the letter "b" ("Foo ^bar"), is this position inside or outside <strong>? If you use native DOM Selection, you may get both positions — one anchored in <p> and the other anchored in <strong>. In CKEditor 5 this position translates exactly to "Foo ^bar".

# Selection attributes

OK, but how to let CKEditor 5 know that I want the selection to “be bold” in the case described above? This is important information because it affects whether or not the typed text will be bold, too.

To handle that, selection also has attributes. If the selection is placed in "Foo ^bar" and it has the attribute bold=true, you know that the user will type bold text.

# Indexes and offsets

However, it has just been said that inside <paragraph> there are two text nodes: "Foo " and "bar". If you know how native DOM Ranges work you might thus ask: “But if the selection is at the boundary of two text nodes, is it anchored in the left one, the right one, or in the containing element?”

This is, indeed, another problem with DOM APIs. Not only can positions outside and inside some element be identical visually but also they can be anchored inside or outside a text node (if the position is at a text node boundary). This all creates extreme complications when implementing editing algorithms.

To avoid such troubles, and to make collaborative editing possible for real, CKEditor 5 uses the concepts of indexes and offsets. Indexes relate to nodes (elements and text nodes) while offsets relate to positions. For example, in the following structure:

    "Foo "

The "Foo " text node is at index 0 in its parent, <image></image> is at index 1 and "bar" is at index 2.

On the other hand, offset x in <paragraph> translates to:

offset position node
0 <paragraph>^Foo <image></image>bar</paragraph> "Foo "
1 <paragraph>F^oo <image></image>bar</paragraph> "Foo "
4 <paragraph>Foo ^<image></image>bar</paragraph> <image>
6 <paragraph>Foo <image></image>b^ar</paragraph> "bar"

# Positions, ranges and selections

The engine also defines three levels of classes which operate on offsets:

  • A Position instance contains an array of offsets (which is called a “path”). See the examples in Position#path API documentation to better understand how paths work.
  • Range contains two positions: start and end ones.
  • Finally, there is a Selection which contains one or more ranges, attributes, and has a direction (whether it was done from left to right or right to left). You can make as many instances of it as you needed and you can freely modify it whenever you want. Additionally, there is a single DocumentSelection. It represents the document’s selection and can only be changed through the model writer. It is automatically updated when the document’s structure is changed.

# Markers

Markers are a special type of ranges.

  • They are managed by MarkerCollection.
  • They can only be created and changed through the model writer.
  • They cane be synchronized over the network with other collaborating clients.
  • They are automatically updated when the document’s structure is changed.
  • They can be converted to attributes or elements in the view.

This makes them ideal for storing and maintaining additional data in the model – such as comments, selections of other users, etc.

# Schema

The model’s schema defines several aspects of how the model should look:

  • where a node is allowed or disallowed (e.g. paragraph is allowed in $root, but not in heading1),
  • what attributes are allowed on a certain node (e.g. image can have src and alt attributes),
  • additional semantics of model nodes (e.g. image is of the “object” type and paragraph of the “block” type).

This information is then used by features and the engine to make decisions how to process the model. For instance, the information from the schema will affect:

  • what happens with a pasted content - what is filtered out (note: in case of pasting the other important mechanism is the conversion – HTML elements and attributes which are not upcasted by any of the registered converters are filtered out before they even become model nodes, so the schema is not applied to them; we’ll cover conversion later in this guide),
  • to which elements the heading feature can be applied (which blocks can be turned to headings and which elements are blocks in the first place),
  • which elements can be wrapped with a block quote,
  • whether the bold button is enabled when the selection is in a heading (and whether text in this heading can be bolded),
  • where the selection can be placed (which is – only in text nodes and on object elements),
  • etc.

The schema is, by default, configured by editor plugins. It is recommended that every editor feature come with rules which enable and preconfigure it in the editor. This will make sure that the plugin’s user can enable it without worrying to re-configure her/his schema.

Currently, there is no straightforward way to override the schema preconfigured by features. If you want to override the default settings when initializing the editor, the best solution is to replace editor.model.schema with a new instance of. This, however, requires rebuilding the editor.

The instance of the schema is available in editor.model.schema. Read an extensive guide about using the schema API in API docs for the Schema class.

# View

Let’s again take a look at the editing engine’s architecture:

Diagram of the engine’s MVC architecture.

So far, we talked about the topmost layer of this diagram – the model. The role of the model layer is to create an abstraction over the data. Its format was designed to allow storing and modifying the data in the most convenient way, while enabling implementation of complex features. Most features operate (read from it and change it) on the model.

The view, on the other hand, is an abstract representation of the DOM structure which should be presented to the user (for editing) and which should (in most cases) represent the editor’s input/output (i.e. data returned by editor.getData(), data set by editor.setData(), pasted content, etc.).

What this means is that:

  • The view is yet another custom structure.
  • It resembles the DOM. While the model’s tree structure only slightly resembled the DOM (e.g. by introducing text attributes), the view is much closer to the DOM. In other words, it is a virtual DOM.
  • There are two “pipelines” – the editing pipeline (also called the “editing view”) and the data pipeline (“the data view”). Treat them as two separate views of one model. The editing pipeline renders and handles the DOM which the user sees and can edit. The data pipeline is used when you call editor.getData(), editor.setData() or paste content into the editor.
  • The views are rendered to the DOM by the Renderer which handles itself all the quirks required to tame the contentEditable used in the editing pipeline.

The fact that there are two views is visible in the API:

editor.editing;                 // The editing pipeline (EditingController).
editor.editing.view;            // Editing view's controller.
editor.editing.view.document;   // Editing view's document.;                    // The data pipeline (DataController).

Technically, the data pipeline does not have a document and a view controller. It operates on detached view structures, created for the purposes of processing a data.

It is much simpler than the editing pipeline and in the following part of this section we will be talking about the editing view.

Check out the EditingController's and DataController's API.

# Changing the view

Do not change the view manually, unless you really know what you do. If the view needs to be changed, in most cases, it means that the model should be changed first. Then, the changes you apply to the model are converted (conversion is covered below) to the view by specific converters.

The view may need to be changed manually if the cause of such a change is not represented in the model. For example, the model does not store information about the focus, which is a property of the view. When the focus changes, and you want to represent that in some element’s class, you need to change that class manually.

For that, just like in the model, you should use the change() block (of the view) in which you will have access to the view writer. writer => {
    writer.insert( position1, writer.createText( 'foo' ) );
} );

# Element types and custom data

The structure of the view resembles the structure in the DOM very closely. The semantics of HTML is defined in its specification. The view structure comes “DTD-free”, so in order to provide additional information and better express the semantics of the content, the view structure implements 5 element types (ContainerElement, AttributeElement, EmptyElement, UIElement, and EditableElement) and so called “custom properties” (i.e. custom element properties which are not rendered). This additional information provided by editor features is then used by the Renderer and converters.

The element types can be defined as follows:

  • Container element – elements which build the structure of the content. Used for block elements such as <p>, <h1>, <blockQuote>, <li>, etc.
  • Attribute element – elements which cannot contain container elements inside them. Most model text attributes are converted to view attribute elements. They are used mostly for inline styling elements such as <strong>, <i>, <a>, <code>. Similar attribute elements are flattened by the view writer, so e.g. <a href="..."><a class="bar">x</a></a> would automatically be optimized to <a href="..." class="bar">x</a>.
  • Empty element – elements which must not have any child nodes – e.g. <img>.
  • UI elements – elements which are not part of the “data” but needs to be “inlined” in the content. They are ignored by the selection (it jumps over them) and the view writer in general. Content of these elements and events coming from them are filtered out too.
  • Editable element – elements used as “nested editables” of non-editable fragments of the content – e.g. caption in the image widget, where the <figure> wrapping the image is not editable (it is a widget) and the <figcaption> inside it is an editable element.

The custom properties are used to store information like:

  • Whether element is a widget (added by toWidget()),
  • How an element should be marked when a marker highlights it,
  • Whether element belongs to a certain feature - it’s a link, progress bar, caption, etc.

# Positions

Just like in the model, in the view there are 3 levels of classes which describe points in the view structure: positions, ranges and selections. A position is a single point in a document. A range consists of two positions (start and end). And selection consists of one or more ranges and has a direction (whether it was done from the left to right or right to left).

A view range is very similar to its DOM counterpart as view positions are represented by a parent and an offset in that parent. This means that, unlike model offsets, view offsets describe:

  • points between child nodes of the position’s parent if it is an element,
  • or points between character of a text node if position’s parent is a text node.

Therefore, we can say that view offsets work more like model indexes than model offsets.

parent offset position
<p> 0 <p>^Foo<img></img>bar</p>
<p> 1 <p>Foo^<img></img>bar</p>
<p> 2 <p>Foo<img></img>^bar</p>
<img> 0 <p>Foo<img>^</img>bar</p>
Foo 1 <p>F^oo<img></img>bar</p>
Foo 3 <p>Foo^<img></img>bar</p>

As you can see, two of these positions represent what you can consider the same point in the document:

  • { parent: paragraphElement, offset: 1 }
  • { parent: fooTextNode, offset: 3 }

Some browsers (Safari, Chrome and Opera) consider them identical too (when used in a selection) and often normalize the first position (anchored in an element) to a position anchored in a text node (the second position). Do not be surprised that the view selection is not directly when you would like it to be. The good news is that the CKEditor 5’s renderer can tell that two positions are identical and avoids re-rendering the DOM selection unnecessarily.

You may sometimes find in the documentation that positions are marked in HTML with the {} and [] characters. The difference between them is that the former indicate positions anchored in text nodes and the latter in elements. So, for example, the following example:


describes a range which starts in the text node Foo at offset 0 and ends in the <p> element at offset 1.

The far-from-convenient representation of DOM positions is yet one more reason to think about and work with model positions.

# Observers

In order to create a safer and more useful abstraction over native DOM events, the view implements the concept of observers. It improves the testability of the editor as well as simplifies the listeners added by editor features by transforming the native events into a more useful form.

An observer listens to one or more DOM events, do preliminary processing of this event and then fires a custom event on the view document. An observer not only creates an abstraction on the event itself but also on its data. Ideally, an event’s consumer should not have any access to the native DOM.

By default, the view adds the following observers:

Additionally, some features add their own observers. For instance, the clipboard feature adds ClipboardObserver.

For the complete list of events fired by observes check the Document's list of events.

You can add your own observer (which should be a subclass of Observer) by using the view.addObserver() method. Check the code of existing observers to learn how to write them:

Since all events are by default fired on Document it is recommended that 3rd party packages prefix their events with an identifier of the project to avoid name collisions. For example, MyApp’s features should fire myApp:keydown instead of keydown.

# Conversion

So far, we talked about the model and the view as about two completely independent subsystems. It is time to connect them. The three main situations in which these two layers meet are:

conversion name description
data upcasting Loading data to the editor.
First, the data (e.g. an HTML string) is processed by a DataProcessor to a view DocumentFragment. Then, this view document fragment is converted to a model document fragment. Finally, the model document’s root is filled with this content.
data downcasting Retrieving data from the editor.
First, the content of model’s root is converted to a view document fragment. Then, this view document fragment is processed by a data processor to a target data format.
editing downcasting Rendering editor’s content to the user for editing.
This process takes place for the entire time when the editor is initialized. First, the model’s root is converted to a view’s root once data upcasting finishes. Then, this view root is rendered to the user in the editor’s contentEditable DOM element (also called “the editable element”). Then, every time the model changes, those changes are converted to changes in the view. Finally, the view can be re-rendered to the DOM if needed (if the DOM differs from the view).

Let’s take a look at the diagram of the engine’s MVC architecture and see where each of the conversion processes happen in it:

Diagram of the engine’s MVC architecture.

# Data pipeline

Data upcasting is a process which starts in the bottom right corner of the diagram (in the view layer), passes from the data view, through a converter (green box) in the controller layer to a model document in the right top corner. As you can see, it goes from the bottom to the top, hence “upcasting”. Also, it is handled by the data pipeline (right branch of the diagram), hence “data upcasting”. Note: data upcasting is also used to process pasted content (which is similar to loading a data).

Data downcasting is an opposite process to data upcasting. It starts in the right top corner and goes down to the right bottom corner. Again, the name of the conversion process matches the direction and the pipeline.

# Editing pipeline

Editing downcasting is a bit different process than the other two.

  • It takes place in the “editing pipeline” (left branch of the diagram).
  • It does not have its counterpart – there is no editing upcasting because all user actions are handled by editor features by listening to view events, analyzing what happened and applying necessary changes to the model. Hence, this process does not involve conversion.
  • Unlike DataController (which handles the data pipeline), the EditingController maintains a single instance of the Document view document’s for its entire life. Every change in the model is converted to changes in that view so changes in that view can then be rendered to the DOM (if needed – i.e. if the DOM actually differs from the view at this stage).

TODO: upcasting, downcasting, mapping nodes and positions, API.

# Architecture of a typical feature

Once you learnt how to implement editing features, it is time to add a UI for them. You can read about the CKEditor 5’s standard UI framework and UI library in the UI library guide.