Facts
U-Discover uses Facts mapped to scorecard questions to extract key information from conversations. It displays the list of Facts along their details.
What is a Fact?
Facts are atomic building blocks of information, each of which can then be used either individually or together to provide insights into conversations. A Fact uses either an LLM or a semantic match to answer questions based on the transcript of a single call. No additional data is used in the evaluation.
A Fact query can be a simple question—“Did the customer inquire about applying for a loan?” or a slightly more detailed query—“Analyze the customer call to determine if the customer is at risk for churn. Specifically, look for signs such as negative feedback, repeated complaints, or expressed dissatisfaction with our services. If the customer is at risk, identify the reasons. Respond with Yes | No : [reason_1, reason_2]”
Facts have different output/response types:
Boolean: Boolean Facts output a “Yes” or “No”. For example, “Did the agent exhibit empathy in the call?”, or “Did the customer request a card cancellation?”. Note: Boolean Facts are used in scorecards. Boolean Facts are supported by both LLM and a semantic match model.
Important
Limitation: When a Fact is created with Fact type as Boolean and query containing random name or number (invalid query) then the Agent’s score is calculated as zero.
Text: Text Facts allow the output of free text. For example, “Did the customer inquire about applying for a loan? What type of loan were they interested in?”. Text Facts are supported only by LLM.
Numeric: Numeric Facts output a numeric value, with or without a unit. For example, “What was the claim amount mentioned?” or “How many times did the customer mention a competitor?”. Numeric Facts are supported only by LLM.
List: List Facts allow the output of a list of pre-configured values. List Facts are supported only by LLM. An option is provided for the LLM to give outputs that are not in the user provided list.
Fact Types
LLM Facts are Facts that are evaluated by a large language model (LLM). The LLM supports all Fact response types – Boolean, Text, Numeric and List.
Semantic Facts are evaluated by a semantic match model, via the use of embeddings. A semantic Fact supports only Boolean output.
A Fact contains the following data:
Name: Name of the Fact
Description: Description of the Fact
Group: A Fact can be assigned to a Group. You can use the Group to filter the list of Facts.
Organization: The Organization is the first grouping under your Account/Tenant. A Fact can be mapped to multiple organizations.
Language: This is the language in which the conversation is spoken. Today only English-US is supported.
Channel: Only Voice is supported.
Direction: This represents which type of call the Fact must be evaluated for – an Inbound call or an Outbound call.
Fact type: This represents the response type of the Fact, whether Boolean, Text, List or Numeric.
Query: If you are creating an LLM-based Fact, this is the most important field associated with the Fact. It is essentially the prompt that is sent to the LLM to evaluate the Fact. You should be as descriptive as possible here. In order to help you fine-tune this query, U-Discover lets you test a Fact as soon as you write the query. The “Test to Validate” feature allows you to test Facts against a test bed of conversations.
Status: Whether Draft or Approved
Testing Facts
System Facts
U-Discover provides certain Facts out of the box categorized as System Facts which is created from the backend. In Conversation Fact page, the System Fact box is checked and the details cannot be modified. Up to 4 System Facts can be created per tenant.
The following is the list of System Facts.
Topic discovery
Agent keyphrases
Customer keyphrases
Summary