How to implement custom applications

Building the mind for your business

How to implement custom applications

Customers and integrators can extend the platform by creating their own client applications and either use them on their account or publish them in their own ecosystem.

Authorization and permission scopes

Applications are exclusively authorized to interact with the system for the purpose of communicating with the dialog engine, meaning that they can only:

Customers willing to perform any other action, for instance injecting data or controlling an account, should instead authenticate with credentials granting the required permissions. See the tutorial on how to create a custom fetcher for more details.

An administrator with permission to write the domain settings can create a new application using the rest APIs. Applications ID (or app_ids) are used by the system to allow access to a client application. The main purpose of an app_id is to identify an app and to revoke its access in case it gets compromised. From a security perspective the app key should be kept secret if possible but it should not considered as a password.

In order to use the REST APIs applications have to request the system for a short lived access token which is given with restricted access acknowledging that the application is recognized but the user using it is not authenticated. This application tokens expire after a short amount of time, when that happens the client is required to get a new one.

Customers are discouraged from re-using the same app_id for multiple purposes because it would reduce the ability to counter malicious attacks. Imagine for example that the same Kare MIND account would be used to power two portal one internal and one public facing with two different applications. If the client application deployed in the public portal is then targeted by a DDoS attack the domain administrator will be able to disable it without impacting the internal portal.

Reference in APIs v2.1:

Getting the default app settings

Administrators with permission to write domain settings can configure the default for how clients should behave. This is normally done from the console but it can also done using the APIs.

Applications have the option to load such configuration or to ignore it. Using the default configuration is useful to allow administrator to remotely control the look and feel of all applications at once. To load the default configuration applications can use the get configuration endpoint.

Reference in APIs v2.1:

Starting a new conversation

As soon a user is ready to start interacting with the dialog engine the application should start a conversation. That can be done by calling the start conversation endpoint.

It is the client responsibility to start new conversations when necessary. This will however affect how messages are aggregated and counted in the rest of the platform. Ideally clients should start a new conversation every time the user is opening the application and start a new one if the user comes back after having closed the app.

Clients are meant to cache the conversation until they need it. It is possible for a client to download all previous messages from a given conversation. However, to prevent malicious attempts of stealing data, downloading historical messages requires permission to read the knowledge which is only granted upon authentication.

Reference in APIs v2.1:

Asynchronous interaction

The recommended way for interacting with the dialog engine is by using a websocket. Once the websocket is connected client and server can exchange messages asynchronously.

curl 'wss://' -H 'Accept-Encoding: gzip, deflate, br' -H 'Accept-Language: en-US,en;q=0.9' -H 'Sec-WebSocket-Key: ' -H 'Upgrade: websocket' -H 'Sec-WebSocket-Extensions: permessage-deflate; client_max_window_bits' -H 'Cache-Control: no-cache' -H 'Connection: Upgrade' -H 'Sec-WebSocket-Version: 13' --compressed

For convenience the WS can also authenticate the client implicitly using the app_id instead of a bearer token.

Messages and communication flow.

The communication flow happens by exchanging messages between the client and the dialog engine. The conversation progress by exchanging messages. Messages have a uri which defines what they are, an aspect which defines how they are meant to be rendered, an id which uniquely identifies them and a parent which identify where they are coming from. Inside a conversation messages are organized as a forest with each query starting a new tree.

The communication flow can be simplified as follows:

  1. The client sends a in/hi to start the conversation.
  2. The engine replies with a out/hi to acknowledge the conversation and with a series of out/* depending ho the configuration.
  3. The client sends a in/query message to submit a question or a in/click to trigger a given action. Sending a query or clicking on a given action are the only two ways with which the user can interact with the engine.
  4. The engine replies with a series of out/*.
  5. The previous two steps are repeated until the user closes the application. If the server decides to terminate the conversation it can do so by sending an out/bye message.

Actions and Commands

Outbound messages can contain trigger actions which the user can activate by clicking them. Actions have triggers which are executed either when the action is shown to the user (on_load) or when the user is clicking on it (on_click).

When triggered actions request the client to execute a certain command. The dialog engine uses commands to describe the desired behaviour to the client, and ideally clients should try to comply with the request to the best of their capabilities.


print instructs the client to print a message in the conversation on behalf of the user. This is intended to give the user a visual aid on what his actions are doing.


Example: print -m="Some text".


send instructs the client to send a message to the server. This is used so that the server can log the action and respond to it.


Example: send -m=<msg>.


close instructs the client to close and terminate. This normally indicates that the conversation has been escalated or is ended.

Example: close.


open instructs the widget to open a resource.


Example: open -href=""


update changes the property of an action.


Example: update -context="warning".


The dialog engine assumes that each client will have an icon set of its choice according with the desired look and feel. When an action requires an icon the engine will suggest which icon to use using an identifier.

The dialog requires the following icons:

out/text aspects

When receiving a out/text message clients are supposed to be able to render the following aspects.

Advanced tracking

The protocol facilitates the collection of signals to better understand the user and to monitor their behaviour.

User tracking

It is possible to trace the user across multiple conversations by passing the user_id url parameter when establishing the connection.


The engine automatically logs as metadata all url parameters starting with the kare-meta- prefix as indicated in the JavascriptSDK documentation.

Tracking custom events

Clients can track their own events by sending in/click messages with the right aspect. The following is an example with a open/link aspect.

When creating a custom event it is important to assign the parent_id of the message used to show the content to the user. The client can assign the new events an id as long as it is a valid uuid, otherwise the engine will auto assign one.

      "text":"Kare developer portal"

Using the REST interface

It is also possible to interact with the dialog engine synchronously by using the REST APIs instead of the websocket. The endpoints are synchronous and wait for the dialog to have published all the messages before returning them.

Customers are advised to use the websocket if possible.

Submitting a query

Applications can submit a query by using the post query endpoint. This has the same effect of sending and in/query message.

The response is the list of messages that the dialog engine has generated to answer the question.

Reference in APIs v2.1:

Submitting a click event

Applications can submit click events by using the post event endpoint. This has the same effect of sending and in/click message.

The response is the list of messages that the dialog engine has generated to continue the conversation.

Reference in APIs v2.1: