Skill - Knowledge Base

Overview

This Knowledge Base Skill gives the Digital Person the ability to answer end user questions based on a customer’s knowledge base. The Knowledge base Skill data is provided in a JSON file with expected questions from the users and corresponding responses. If there is a match the Digital person will pick one of the responses from the list randomly.

This skill is:

• Built and owned by Soul Machines Ltd.

• Only supported in English.

Limitations

1. The Skill does not validate the JSON file. So an invalid JSON file can result in the Skill or project not working as expected.

2. Knowledge Base Skill does not work with other fallback skills like Elegant Failure and Skill - Site Search. However, Elegant Failure is supported as part of the Knowledge Base Skill.

Skill Demo

JSON File Format

The JSON file must be in its raw form or hosted on any static site.

1. It needs to be publicly accessible (no login or other steps)

2. It needs to directly serve the .json file as raw text (not formatted to look nice, not served up in a viewer)

The knowledge base data is a JSON file containing an array of questions and answers. The format of the data is as follows:

1. "intent" - Used as a unique identifier for each JSON object.

2. "examples" - This is an array of strings that contain examples of different phrasings for the question.

3. "metadata" - This contains an array of strings that are the responses to the question. If more than one value is found for a given intent, the skill will randomly pick one of the responses to reply with.

[
  {
    "intent": "WhatIsSoulMachines",
    "examples": [
      "what is soul machines",
      "tell me about soul machines",
      "who is soul machines",
      "who are soul machines",
      "what are soul machines",
      "what is soul machine",
      "who is soul machine",
      "who are soul machine",
      "what are soul machine",
      "tell me about soul machine",
      "explain soul machines to me"
    ],
    "metadata": [
      "Soul Machines is the world leader in humanizing AI to create astonishing Digital People. A little secret, they actually created me.",
      "Soul Machines is my creator"
    ]
  },
  {
    "intent": "HowAreYou",
    "examples": [
      "Hoe are you",
      "thow are you doing",
      "are you doing good today",
      "how are you feeling",
      "how are you going",
      "how are u",
      "how are you going today",
      "how do you feel today",
      "how do you do"
    ],
    "metadata": [
      "I'm good",
      "Great!"
    ]
  }
]

If there is not an exact match, a Natural Language Understanding (NLU) process will attempt to find the closest intent based on the examples. If no match is found, it will reply back with a configurable No Match Response.

Displaying Content cards

In the metadata section of the JSON file, you can either add an array of strings (what the Digital Person will speak when the request comes in), or an array of objects which contain both the spoken text as well as any variables for cards that may be needed. The format of the enhanced metadata response looks as below.

Note that all “variables” should contain a “public” entry under it. Below the “public” level, you can include any number of additional variables, then refer to the name (“pic” in this example). This enhancement supports cards and commands now. See Displaying Content Cards for the list of available cards and their corresponding variable requirements.

{
  "intent": "ShowImage",
  "examples": [
    "show me an image"  ],
  "metadata": [
    {
      "text": "@ShowCards(pic) Here is an image",
      "variables": {
        "public": {
           "pic": {
             "component": "image",
             "data": {
               "url": "https://www.soulmachines.com/wp-content/uploads/SM_NavBarLogo_DarkGray.png"}
          }
        }
      }
    }
  ]
}

Entity support

You can specify “entities” word varieties in the text. In the example below, the entity is in square brackets (i.e. [cats]) the entity can be any word. The main effect of this format for the example is it is more robust in recognizing the pattern and you do not need as many examples to train the NLU.

{
  "intent": "AskAboutPets",
  "examples": [
    "Are [cats] good pets",
    "Are [dogs] good ets",
    "Are [parrots] good pets",
    "Are [hamsters] good pets"  ],
  ],
  "metadata": [
    "All pets are good if they make the owner happy"  ]
}

In the above example, if the user says any other word in place of “cats” the intent is matched and the Digital person will respond with one of the corresponding responses.

Best Practices

1. Aim to add at least 10 phrases in the array for the best possible match for the question

2. Get as many new people to test your Knowledge Base as you can so that you have a wide range of questions and answers. 

3. Make sure your JSON file is validated before you configure the Knowledge Base skill.

Configurations

Provide the following information in the Knowledge Base Skill configuration screen in Digital DNA Studio. See Adding Skills to your Digital Person for detailed instructions: