Autogenerated Knowledge Functionality and the Use of Prompts

This material is intended to help you understand the inner workings of the Autogenerated knowledge functionality and the impact of prompts.

Below, we'll address key points on the topic, which we suggest you read in the order presented to understand the entire journey better.

What is an LLM?

An LLM (Large Language Model) is an artificial intelligence model based on deep learning, trained on vast amounts of text to understand, process, and generate natural language. Its goal is to interact with people organically and solve tasks that previously required human intervention.
How does it work? LLMs use an architecture called Transformer, which allows them to analyze the full context of a text and understand the relationships between words and concepts. They work by predicting the most likely next word in a sequence, generating coherent responses tailored to the context.
Additionally, they can perform new tasks based on instructions (prompts).

What is a prompt in general terms?

"A prompt is the instruction or question you give to an artificial intelligence system to generate a response or perform a task. It’s the ‘input’ you use to guide it, and it can range from a simple phrase to detailed paragraphs.”

How does Autogenerated Knowledge work?

Once you’ve added content to the Autogenerated section of Virtual Assistant—whether from links or PDF documents—and enabled the option to use that Autogenerated content, when one of your customers types a query into the virtual assistant, the process will first search for the information in the Personalized knowledge base; if it doesn’t find what’s being asked, it will then search the Autogenerated knowledge base.

It is at this stage that the guideline specified by the prompt is activated.

The system searches links or documents.
When it finds appropriate content to generate the reference response, it processes it according to the prompt’s instructions.
The information is inserted as context, along with the prompt’s guidelines.
And the response to be delivered to the customer is generated.

Prompt Priority:

The Personalized Knowledge Engine has defined guidelines based on the personality established for the virtual assistant. We’ll refer to this as the Official Prompt.
The default response mode for the virtual assistant is the Official Prompt (the one already set up in the Personalized Knowledge Engine), as it is configured to provide comprehensive and general responses.
If a specific adjustment is needed for a particular area of interest or context, we can add a Custom Prompt.

Important: The custom prompt should only be used to supplement the official prompt. It must not contradict or replace information already included in the official prompt.

Which aspects of the custom prompt can be adjusted?

You will specify the parameters for the voice and tone of the responses provided by the virtual assistant using content autogenerated by Generative AI. You’ll define whether the tone should be formal or friendly, as well as the estimated length of the text the user will receive in the conversation window.

Keep in mind that you must first process the content in the Generative AI knowledge base and enable it.

Role: Define how the virtual assistant should behave (for example, “Official BrandX Virtual Assistant”).
Voice and tone: Set the communication style that best suits the experience you want to offer your users. For example, it can be friendly, formal, approachable, professional, etc.
Estimated maximum number of characters for the response the user will receive: Determine the approximate length of the responses the user will receive in the chat window. We suggest choosing between 200, 300, 400, and 500, keeping in mind that these are estimated character counts.
Text to be used when a non-committal response is necessary: Enter the message the assistant will display when it cannot find information to answer the query or fails to interpret the user’s question. It is recommended that this text not exceed 400 characters.

Additionally, you can add guidelines regarding derivations, integrations, buttons, and other complements. See the Key Points section.

Complements are generated using internal HTML tags that are recognized by our services. In other words, this is not an agent that currently performs tasks, but rather one that responds with tags that our system interprets and, based on those tags, executes an action.

How do you request the parameters to be applied to your customizable prompt for Autogenerated Knowledge?

Submit a request through our Support Center.

Key Points:

If, as a customer, you request these specific additions from the Support team, you should be aware that LLMs are non-deterministic models—that is, they do not always generate the same response to the same input (unlike a traditional system based on fixed rules).
This happens because the model chooses each word based on probabilities, not on a single predefined “correct” answer. Therefore, even if a rule is correctly written in the prompt, there is no guarantee that the LLM will always follow it in the same way, so it may sometimes not respond as expected.
This behavior is further accentuated when adding derivations, integrations, buttons, and other complements, due to the nature of LLMs, which are designed exclusively to respond with text.
It is essential to avoid setting prompts that are too long to ensure that the model's performance is not affected.

What Should You Avoid Doing?

> You should not include prompts that tell the virtual assistant how to solve problems within the documents you add to the Autogenerated Knowledge section.

Why?

This alters the technical behavior of the feature, which negatively impacts your users’ experience when they receive a response.
The true purpose of Autogenerated Knowledge—in its full scope, whether incorporating links or PDF documents—is the massive and systematic extraction of information using software programs or tools, without manual intervention. In other words, it’s about gathering knowledge to feed the virtual assistant, so Generative AI can formulate possible responses users may need.

If you still have questions, please consult your CSM or contact us at our Support Center.