Knowledge Assets Best Practices

Answer Quality Improvement

If a user interacts with an Airline dataset such as the following:

Users may inquire about specific airports using their names (for example, Dulles or Raegan) rather than airport codes (for example, IAD or DCA), as shown below:

In this example, the Bot is unable to interpret the significance or the meaning of the names provided. To remedy this, you can provide detailed information in a single-column Excel file. The file should define the airport codes across three separate rows, as shown in the following image. For example, "IAD is the airport code for Washington Dulles International Airport".

Once the file is uploaded, studied, and saved, as described above, you can receive accurate responses that leverage the knowledge in the uploaded file.

Only the relevant information is passed to MicroStrategy AI.

See the accurate responses from the Bot after uploading the Excel file:

Use Cases and Best Practices

Follow the following best practices when using knowledge assets:

Each knowledge asset must include some keywords or words with similar semantic definitions used in both the question and the knowledge asset. These similarities are required to allow the knowledge asset to be used when answering a question. If the knowledge asset includes generic rules, Auto Answers and Bots will have issues semantically relating knowledge to the question. Generic rules can be placed in the Bot Custom Instructions.

Incorporate definitions that clarify object names for the Bot including terms, acronyms, or synonyms. For example:

The code SLS stands for Sales
The code SOP is a descriptor for SLS - Operations with role definition as Operational support for Sales Activities (AP, SL, ES, SE)
                                                        

Provide definitions for existing objects to guide users to form their questions with the correct terms. For example, when a user specifies a metric such as the following:

L4Q stands for last four quarter and is a metric of average score for last 4 quarters. It can also be considered as the annual score.
                                                        

Defining business jargon is beneficial before incorporating them into analysis:

Include specific definitions that are critical to analyze the data. For example:

For a human male of age 0-16, the normal blood pressure is 116/70mm Hg
For a human male of age 17-35, the normal blood pressure is 119/79mm Hg
For a human male of age 36-59, the normal blood pressure is 124/77mm Hg
For a human male of age 60 and above, the normal blood pressure is 133/69mm Hg
For a human female of age 0-16, the normal blood pressure is 105/66mm Hg
For a human female of age 17-35, the normal blood pressure is 122/72mm Hg
For a human female of age 36-59, the normal blood pressure is 132/70mm Hg
For a human female of age 60 and above, the normal blood pressure is 139/68mm Hg
                                                        

Using the knowledge above, the Bot can provide more accurate results by considering the patient's age from the dataset, as the generalized benchmark of 120/80 mm Hg will not be applied universally across all ages and genders.

Calculate derived metrics using existing ones. For example, where Revenue and C3 are existing metrics:

Predicted Revenue = [0.23 * (Revenue)] + C3

Definitions for attribute element values can be included to enhance the Bot's comprehension of user queries. As explained in Answer Quality Improvement, the knowledge can be defined as the following:

The attribute Airport Code contains values like BWI, IAD and DCA.
BWI is the airport code for Baltimore/Washington International Thurgood Marshall Airport.
DCA is the airport code for Ronald Reagan Washington National Airport.
IAD is the airport code for Washington Dulles International Airport.
                                                        

If you want to test the effectiveness of an additional piece of data before adding it to the knowledge asset, you can add the knowledge to the query itself. For example:

How many SBD employees did we hire this year ? The code SLS stands for Sales.  The code SBD is a descriptor for SLS - Business Development with role definition as Responsible for Business Development Support for Sales.
                                                        

If the additional context returns the expected answer, you can add it to the knowledge asset.

Avoid the following behavior when using knowledge assets:

Definitions should not contain irrelevant information.
Answer performance depends on the your knowledge asset row and the number of relevant rows in the question. Avoid adding too many rows or too much information to each row. This can confuse the AI and result in hallucinations.

For example, if you expect questions based on roles and responsibilities for employees in a sales department, you can use the following definitions:

Copy

The code SLS stands for Sales. The code SOP is a descriptor for SLS - Operations with role definition as Operational support for Sales Activities (AP, SL, ES, SE).The code SBD is a descriptor for SLS - Business Development with role definition as Responsible for Business Development Support for Sales. The code SCS is a descriptor for SLS - Customer Success Manager with role definition as Responsible for Account Health Checks & Renewals.

However, passing the entire text to the MicroStrategy AI increases the token count and delays the response time. Instead, the information should be broken down into multiple rows:

Copy

The code SLS stands for Sales. 
The code SOP is a descriptor for SLS - Operations with role definition as Operational support for Sales Activities (AP, SL, ES, SE).
The code SBD is a descriptor for SLS - Business Development with role definition as Responsible for Business Development Support for Sales. 
The code SCS is a descriptor for SLS - Customer Success Manager with role definition as Responsible for Account Health Checks & Renewals.

Avoid embedding additional information about attributes within the text and expecting the Bot to extract it then use it to filter and aggregate the data.

For example, consider a dataset with Continent, Country, and City attributes and Population and Area metrics.

In this example, if a user asks: "How many cities that I've traveled have French as a spoken language?", the following knowledge has additional information about the Country attribute and can make the AI hallucinate.

Copy

The city Strasbourg in France has a population of 280000. The official spoken language is French.
The city Tours in France has a population of 140000. The official spoken language is French.
The city Berlin in Germany has a population of 3500000. The official spoken language is German.
The city Warsaw in Poland has a population of 1700000. The official spoken language is Polish.
The city Montreal in Canada has population of 1700000. The official spoken language is French.

Upload knowledge in the following form to avoid hallucination:

Copy

Cities such as Strasbourg, Tours, Lyons, Montreal, Abidjan, Yaounde, Madagascar, and Dakar use French as the official spoken language.