Below are best practices to keep in mind when working with content created by Generative AI from the “Autogenerated” section of Virtual Assistant. These will help you optimize its use and ensure that the responses your virtual assistant provides across different customer service channels are helpful and accurate.
Preparing Source Content (PDFs and Links)
- Structure the content: It must follow a logical, well-organized order so Generative AI can process it properly. This involves clear titles and comprehensive descriptions. Titles should provide context for the response and also specify the channel, for example, “How do I change my password in the app?” or “How do I change my password on the website?”
- Distinguish between similar content: It is essential to avoid duplicate content in your Personalized knowledge base to maintain consistency across topics. If there are similar topics in your personalized content, ensure that both the title and the body text are sufficiently distinct so that Generative AI can distinguish them correctly.
- Avoid duplicate information within the same document or across different documents: It is recommended that the content you include in the Autogenerated Knowledge section (whether within the same PDF or across different documents/links) not be repeated. This prevents Generative AI from having to compare or decide which content is more valid or relevant when there is varying information on the same topic.
- Changing information: If you need to include information that is subject to frequent changes, such as prices, it is recommended that you avoid using Generative AI for that content and use Personalized Knowledge instead.
- One language per document or link: One language per document or link: The PDF text or URLs added to the content to be processed by Generative AI must be in the same language as the bot. The same applies to any customized prompts. This improves understanding and response efficiency.
- Character correction and use of optical character recognition (OCR): When training with PDF documents, OCR may confuse characters—for example, mistaking a lowercase “l” for an uppercase “I.” To avoid this, it is recommended to use fonts such as Times New Roman, which are more reliable in these processes.
- Use of links: Links must use a secure browsing protocol (HTTPS), be publicly accessible, and be visible to allow access to the information.
- Tables, charts, and images: Although Generative AI can read these formats, it is recommended that you review the responses it generates to ensure that it has interpreted the information correctly. This ensures that the content is clear, accurate, and actionable, avoiding potential conceptual errors in the automated responses. If necessary, convert the information into descriptive text for greater reliability.
- You should not include prompts that instruct the virtual assistant on how to process the documents you add to the Autogenerated Knowledge section. Why? This modifies the technical behavior and can hurt the user experience. 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 or computational tools, without manual intervention. In other words, it’s about obtaining knowledge to feed the virtual assistant so that Generative AI can formulate potential responses users may need.
Custom Prompt
- Small but effective adjustments: Minor tweaks to the custom prompt can improve the effectiveness of the responses. This includes specifying the format, response length, or other details to ensure that the Generative AI response is more accurate and consistent with what is expected.
- Maintain clarity and objectivity: When creating a prompt, it’s important to be clear and objective. This makes it easier for Generative AI to interpret the prompt, leading to more relevant and focused responses. Avoid excessive or irrelevant information.
- Adapting to the context: If the situation or query requires a more specialized response, the custom prompt can be adjusted to reflect the interaction's context better, making it more focused.
- Additional instructions: Include simple instructions in the content, such as “speak with an advisor,” if the user needs further clarification or human support. This guides the user toward the appropriate assistance when Generative AI cannot resolve the query.
- The custom prompt must not conflict with the official prompt and is meant to enhance the quality of the response. The goal is to improve the quality of the responses, including their length, accuracy, and correctness.
 Please review this material for more information, as this configuration can only be performed by our Support team:Autogenerated Knowledge Functionality and the Use of Prompts |
Assistant Architecture and Configuration
- Using Generative AI and Multibots: If you use Generative AI to rewrite content, we recommend doing so in a separate bot, especially if you have the multibot option. This will help ensure the main bot is not affected by errors that may occur during content optimization. The same principle applies to training with educational content or academic topics: we recommend using one bot per subject to avoid confusion, since students often do not provide enough context in their queries; therefore, dividing the topics will improve the bot’s accuracy.
- Avoid sensitive information in Generative AI: Since Generative AI can make errors, it is highly recommended that any information that may constitute sensitive or critical data be carefully trained in the custom engine using the appropriate review and approval process.
- Fixed, Unique Responses: If you want responses to be unique and fixed—that is, for the bot to provide the same response every time it is queried—it is recommended that you use Personalized Knowledge rather than Generative AI.
Testing, Monitoring, and Response Quality
- Preproduction testing: Test the generated responses in a controlled environment before deploying them to production to identify potential errors or necessary adjustments.
- Specific end-user questions: End-user questions should be, just as with the standard engine, concrete and specific. Asking a question with just one word is not the best way to get an answer, as you might get no answer, an answer that is vague, or the wrong answer. The question, as with any bot, should be a complete sentence.
- Ongoing review: Frequently review the content auto-generated by Generative AI to ensure it remains aligned with your company’s policy and does not deviate from key messages. It's always best to keep testing and improving things. Use different types of questions to see how well the prompt works and make changes as needed. This helps identify weaknesses and opportunities for improvement to ensure consistency in the generated responses.
- Clarity and accuracy: Ensure that the generated responses are clear and specific, avoiding ambiguities or misconceptions.
If you still have questions, please consult your CSM or contact us at our Support Center.