spiegel-keyman/common/predictive-text/docs/worker-communication-protocol.md

Keyman/Language Modelling layer

This document introduces the protocol for communicating between KeymanWeb and the language modeling layer (i.e., the configurable prediction and suggestion engine, (henceforth referred to as the "LMLayer").

Note: I'm using the term keyboard as a synonym for KeymanWeb.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

Communication protocol between keyboard and asynchronous worker

We have decided that everything to the right of the KeymanWeb will be in a Web Worker. However, communication can happen only through postMessage(data) commands, where data is a serializable object (via the structured clone algorithm).

What serializable object can we send that will adhere to the open-closed principle?

Messages

The idea is to use a discriminated union. The protocol involves plain JavaScript objects with one property called message that takes a finite set of string values.

These string values indicate what message should be sent. The rest of the properties in the object are the parameters send with the message.

{
    message: 'predict',
    // message-specific properties here
}

See also: XML-RPC

Messages are not methods. That is, there is no assumption that a client will receive a reply when a message is sent. However, some pairs of messages, such as predict and suggestions, assume (non-blocking!) request/response semantics.

Tokens

Tokens uniquely identify an input event, such as a key press. Since Keyman should ask for an asynchronous prediction on most key presses, the token is intended to associate a prediction and its response with a particular input; Keyman is free to ignore prediction responses if they are for outdated input events.

The Token type is opaque to LMLayer. That is, LMLayer does not inspect its contents; it simply uses it to identify a request and pass it back to Keyman. There are a few requirements on the concrete type of the Token:

1. Tokens MUST be serializable via the structured clone algorithm;
2. Tokens MUST be usable as a key in a Map object.
3. Tokens MUST be unique across messages. That is, tokens MUST NOT be duplicated between different messages.

It is up to the keyboard to create unambiguous tokens that can be uniquely identified through the round-trip process.

In the following examples, I'll use the subset of number values that are interpretable as 31-bit signed integers.

Example

An asynchronous message to predict after typing 'D':

{
    message: 'predict',

    token: 1,
    transform: {
        insert: 'D',
        deleteLeft: 0,
        deleteRight: 0
    },
    context: {
      left: '',
      right: '',
      startOfBuffer: true,
      endOfBuffer: true
    },
}

Message types

Currently there are four message types:

Message	Direction	Parameters	Expected reply	Uses token
config	LMLayer → worker	capabilities	No	No
load	keyboard → LMLayer	model	Yes — ready	No
unload	keyboard → LMLayer	none	No	No
ready	LMLayer → keyboard	configuration	No	No
predict	keyboard → LMLayer	transform, context	Yes — suggestions	Yes
suggestions	LMLayer → keyboard	suggestions	No	Yes
wordbreak	LMLayer → worker	context	Yes - currentword	Yes
currentword	LMLayer → keyboard	string	No	Yes

Message: config

Must be sent from the keyboard to the LMLayer so that the LMLayer may properly configure loaded models. It will send config, a plain JavaScript object specifying platform restrictions.

The keyboard MUST NOT send any messages to the LMLayer prior to sending config. After this, it is safe to assume the config was performed successfully and is ready to load a model.

The LMLayer needs to know the platform's abilities and restrictions (capabilities).

interface LoadMessage {
  message: 'load';

  /**
   * The path to the model's compiled script file.
   */
  capabilities: {
    /**
     * The maximum amount of UTF-16 code units that the keyboard will provide to
     * the left of the cursor, as an integer.
     */
    maxLeftContextCodeUnits: number,

    /**
     * The maximum amount of code units that the keyboard will provide to the
     * right of the cursor, as an integer. The value 0 or the absence of this
     * rule implies that the right contexts are not supported.
     */
    maxRightContextCodeUnits?: number,

    /**
     * Whether the platform supports deleting to the right. The absence of this
     * rule implies false.
     */
    supportsDeleteRight?: false,
  }
}

Message: load

Must be sent from the keyboard to the LMLayer so that the LMLayer loads a model. It will send load which is a plain JavaScript object specify the path to the model, as well configurations and platform restrictions.

After a single config message, the keyboard MUST NOT send any messages to the LMLayer prior to sending load. The keyboard SHOULD NOT send another message to the keyboard until it receives ready message from the LMLayer before sending another message.

The LMLayer needs to know which concrete language model to instantiate. This is provided by the file at the path specified by the model string parameter.

interface LoadMessage {
  message: 'load';

  /**
   * The path to the model's compiled script file.
   */
  model: string
}

Message: unload

Must be sent from the keyboard to the LMLayer so that the LMLayer resets itself in preparation for loading a new model. It will send unload which is a plain message to trigger release of old model resources.

interface UnloadMessage {
  message: 'unload';
}

Message: ready

Must be sent from the LMLayer to the keyboard when the LMLayer's model as a response to load. It will send configuration, which is a plain JavaScript object requesting configuration from the keyboard.

There are only two options defined so far:

interface ReadyMessage {
  message: 'ready';
  configuration: {
    /**
     * How many UTF-16 code units maximum to send as the context to the
     * left of the cursor ("left" in the Unicode character stream).
     *
     * Affects the `context` property sent in `predict` messages.
     *
     * While the left context MUST NOT bisect surrogate pairs, they MAY
     * bisect graphical clusters.
     */
    leftContextCodeUnits: number,

    /**
     * How many UTF-16 code units maximum to send as the context to the
     * right of the cursor ("right" in the Unicode character stream).
     *
     * Affects the `context` property sent in `predict` messages.
     *
     * While the left context MUST NOT bisect surrogate pairs, they MAY
     * bisect graphical clusters.
     */
    rightContextCodeUnits: number,
  };
}

Message: predict

Sent from the keyboard to the LMLayer whenever a new prediction should be generated. This is typically initiated by a key press event. The keyboard SHOULD track each predict message using a token. The token MUST be unique across all prediction events. The LMLayer SHOULD respond to each predict message with a suggestions message. The suggestions message MUST contain the corresponding token as sent in the initial predict message.

The keyboard MUST send the context parameter. The keyboard SHOULD send the transform parameter. The keyboard MUST send a unique token.

The semantics of the predict message MUST be from the perspective of this sequence of events:

1. After the input event is received by the keyboard.
2. Before the keyboard applies the associated transform to the buffer.

NOTE: The keyboard MAY apply the transform associated with the input event before receiving the corresponding suggestions message from the LMLayer. The intention is that once the suggestions are displayed, the typist may select one of the suggestions in the place of the effects of their original input.

NOTE: The keyboard MAY send the predict message after it has applied the transform associated with the input event to the buffer. Regardless of the actual sequence, the semantics MUST remain the same: the prediction happens from the perspective before the transform has been applied. Therefore, if the keyboard eagerly transforms the buffer before it has sent the predict message, it must anticipate undoing the transform it has already applied should it apply a transfrom sent in the suggestions message. The keyboard must act as if it has never applied the transform associated with the input event in the first place.

Context

The context is the text surrounding the insertion point, before the transform is applied to the buffer.

interface Context {
  /**
   * Up to maxLeftContextCodeUnits code units of Unicode scalar value
   * (i. e., characters) to the left of the insertion point in the
   * buffer. If there is nothing to the left of the buffer, this returns
   * an empty string.
   */
  left: USVString;

  /**
   * Up to maxRightContextCodeUnits code units of Unicode scalar value
   * (i. e., characters) to the right of the insertion point in the
   * buffer. If there is nothing to the right of the buffer, this returns
   * an empty string.
   */
  right?: USVString;

  /**
   * Whether the insertion point is at the start of the buffer.
   */
  startOfBuffer: boolean;

  /**
   * Whether the insertion point is at the end of the buffer.
   */
  endOfBuffer: boolean;
}

Transform

The transform parameter describes how the input event will change the buffer.

interface Transform {
  /**
   * The Unicode scalar values (i.e., characters) to be inserted at the
   * cursor position.
   *
   * Corresponds to `s` in com.keyman.KeyboardInterface.output.
   */
  insert: USVString;

  /**
   * The number of code units to delete to the left of the cursor.
   *
   * Corresponds to `dn` in com.keyman.KeyboardInterface.output.
   */
  deleteLeft: number;

  /**
   * The number of code units to delete to the right of the cursor.
   * Not available on all platforms.
   */
  deleteRight?: number;
}

Message: suggestions

The suggestions message is sent from the LMLayer to the keyboard. This message sends a ranked array of suggestions, in descending order of probability (i.e., entry 0 is most likely, followed by entry 1, etc.). This message MUST be in response to a predict message, and it MUST respond with the corresponding token.

/**
 * `suggestions` is an ordered array of suggestion objects.
 * Each suggestion is a transform bundled with a `displayAs` property.
 */
let suggestions = Suggestion[];

interface Suggestion {
  /**
   * Same object as an input event transform.
   * Note that the transform is applied AFTER the input event
   * transform.
   */
  transform: Transform;
  /**
   * A string to display the suggestion to the typist.
   * This should aid the typist understand what the transform
   * will do to their text.
   */
  displayAs: string;
}

let suggestions = [
  {
    transform: {
      insert: 'teapot',
      deleteLeft: 1,
      deleteRight: 0
    },
    displayAs: '🍵'
  }
];

Timing

Each suggestion provides a transform. This transform is applied after the transform associated with the input event that initiated this prediction. That is, the suggested transform applies to the buffer after the transform associated with the input event.

Rephrased in somewhat mathematical terms: Let 𝑥 ∈ 𝐵 be the input text buffer. Let 𝑇_𝑖 be the transform associated with an input event that maps a text buffer 𝐵 → 𝐵. Let 𝑇_𝑠 be a transform suggested through the LMLayer. 𝑦 ∈ 𝐵 is the text buffer after the suggestion transform has been applied. The correct sequence of applications should be as follows:

𝑦 = 𝑇_𝑠(𝑇_𝑖(𝑥))

Late suggestions

Sometimes, a suggestions message may arrive after an input event has already invalidated its request. This is called a "late" suggestion. The LMLayer MAY send late suggestions. Consequently, the keyboard MAY discard the late suggestions. There is no requirement for the keyboard to acknowledge late suggestions, or for the LMLayer to avoid sending late suggestions messages. In either case, a suggestions message can be identified as appropriate or "late" via its token property.

Message: wordbreak

Sent from the keyboard to the LMLayer to request a wordbreak operation on the current context for use in its UI operations. (One useful example: display text for a 'reversion' suggestion that undoes previously-applied suggestions.) The keyboard SHOULD track each wordbreak message using a token. The token MUST be unique across all wordbreak events. The LMLayer SHOULD respond to each wordbreak message with a currentword message. The wordbreak message MUST contain the corresponding token as sent in the initial wordbreak message.

The keyboard MUST send the context parameter. The keyboard MUST send a unique token.

The semantics of the wordbreak message MUST be from the perspective of this sequence of events:

1. After any pending UI operations are completed, including application of any selected suggestion's transform to the context.
2. Before the user inputs any additional keystrokes.

Message: currentword

The currentword message is sent from the LMLayer to the keyboard. This message sends the single word (as determined by wordbreaking) from within context that begins to the left of the insertion point but following the last wordbreaking character(s) to the left of the insertion point. If no such wordform exists, this returns an empty string.

This message MUST be in response to a wordbreak message, and it MUST respond with the corresponding token.

Examples

{
    message: 'wordbreak',

    token: 1,
    context: {
      left: 'The quick brown fox jumped',
      right: ' over the lazy dog',
      startOfBuffer: true,
      endOfBuffer: true
    },
}

would result in the following reply message:

{
    message: 'wordbreak',

    token: 1,
    word: 'jumped'
}

Alternatively, if the insertion point is after the space...

{
    message: 'wordbreak',

    token: 1,
    context: {
      left: 'The quick brown fox jumped ',
      right: 'over the lazy dog',
      startOfBuffer: true,
      endOfBuffer: true
    },
}

would result in the following reply message:

{
    message: 'wordbreak',

    token: 1,
    word: ''
}

The latter example returns empty string as the final pre-insertion point word preceeds the final pre-insertion point wordbreaking character.

Late responses

Sometimes, a currentword message may arrive after an input event or other UI interaction has already invalidated its request. The LMLayer MAY send late results. Consequently, the keyboard MAY discard the late wordbreaking currentword result. This is no requirement for the keyboard to acknowledge late wordbreaking results, or for the LMLayer to avoid sending such late messages. In either case, a currentword message can be identified as appropriate or "late" via its token property.

Informative: LMLayer worker as a state machine

The LMLayer worker can be seen in the following states:

• unconfigured
• model-less
• ready

The transitions of this diagram correspond to messages as described above.