Table of Contents

How to Write the Best Knowledge Sources for AI Agents

Sana Updated by Sana

Knowledge Sources (KS) provide essential information to your AI Agents, helping them engage effectively with customers. While AI Agents have basic capabilities, customizing them with business-specific data is crucial for handling specialized conversations autonomously.

In this article, you will learn about the best knowledge sources you can provide your AI Agents.

What Business-Specific KS do AI Agents Need?

Business-specific information includes your business policies, processes, product pricing, specifications, and more. Respond AI doesn't have knowledge of this information by default. You need to provide it to your AI Agents so they can effectively engage with customers.

What to Use as a Knowledge Source?

You can use your existing business knowledge base to educate your AI Agents or even create new sources. Here are some examples:

  • Manuals, Guides or FAQs (Recommended): Manuals, Guides, Help Center and FAQ documents are very good information sources and are highly recommended as they are:
    • Well-structured.
    • Information is easy to process and absorb.
    • Easy to refer to later and answer Contact queries efficiently.
      For example, FAQs are designed to address common questions and concerns that Contacts have about your business. By incorporating FAQ content as a Knowledge Source, your AI Agent can handle simplistic conversations autonomously.
  • Homepage/Websites/Blogs: Business homepages or product websites are a good source of latest information about your products. You can use the auto-sync feature to automatically sync the knowledge source with the website at set intervals.
    • AI Agents are made aware of the latest promotions, offers and other information about your business.
    Web pages with too many images or poorly structured content are not ideal sources of information. Such webpages don't make the best Knowledge Sources.
  • Conversation Transcripts: Existing conversation transcripts can provide valuable insights and information to your AI Agents regarding how to engage in natural and meaningful conversations with Contacts.
    • As a result, human and AI Agents conduct conversations in a similar manner.
    • Real conversations often contain domain-specific information that other KS might not have. Lack of context can make it hard for the AI Agent to understand when to use the information while conversing with the Contacts.
    The AI Agents must be provided specific information about the domain/industry they are expected to operate in.
    • Real interactions also provide language cues to AI Agents. For example, the AI Agent can learn how to better handle slang language and varying speech patterns.
    Conversation transcripts must be verified for accuracy, and any information that should not be made public must be removed before.

Writing Recommendations

Here are some recommendations that you can keep in mind while you develop your Knowledge Sources.

  • Clear and concise content that is easy to understand and convey to Contacts. Avoid jargon and complex language that make the information difficult to process.
  • Include information that addresses common queries, concerns, and pain points, as well as any unique characteristics or preferences of the target audience.
  • Maintain separate knowledge sources for separate languages.
  • Keep maintaining KS by providing it with the latest information.
  • Rewrite KS to improve readability for AI if your existing KS does not meet the standards recommended in this article.
  • Analyse conversations between Contacts and AI Agents. Identify where the AI Agent has hallucinated and find out the probable cause.
The probable cause can either be that the AI Persona needs to be tweaked or the KS is not sufficient. Learn more about AI Persona best practices here.
  • Test your solutions and eliminate the problem. Continue testing until your AI has been optimized for your needs.

Localization Recommendations

  • Use English as the Primary Language: Based on the latest findings, Knowledge Sources perform best when written in English. This ensures more accurate and reliable responses, as AI models like ChatGPT are primarily trained in English.
  • Multi-Language Setup: Instead of creating separate Knowledge Sources for each language, keep your Knowledge Sources in English and configure the AI Agent to respond in the necessary language. This approach minimizes inconsistencies and errors, particularly with complex inquiries.

Writing Best Practices

  • Focused and precise info: Focused information helps the AI Agent find information easily. Ensure that the information is as precise as possible
  • Task-Based Knowledge Organization: Structure information or data based on specific tasks or functions the AI Agent is expected to handle.
    • First, use AI Persona to inform the AI Agent of their role and then provide them with knowledge sources specific to that role.
    • Then, use KS to equip the AI with the knowledge it needs to effectively communicate with Contacts.
  • Use Headings: Headings create a hierarchy and enhance the readability of the information for both humans and AI systems.
  • Indent Info: Indents also create a hierarchy and make the content easier to scan through.
  • Use bold to highlight: Highlight important information by setting the text as bold.
  • Describe images: Our current AI model (GPT-4o mini) cannot process images. However, as a workaround, you can describe the images in the knowledge source.
    • Example of how to describe an image:
      • This is a bar chart showing quarterly sales figures with bars labeled Q1 through Q4, and sales amounts ranging from $10,000 to $50,000.
    Learn more about converting images to acceptable formats here.
  • Describe tabular data: Our current AI model (GPT-4o mini) cannot process data in tables. However, as a workaround, you can describe the tables in the knowledge source.
    • Example of how to describe a table:
      • Vegetables available in stock:
        • Apple:
          Availability: 20 in stock; Price: 1$
        • Pear:
          Availability: 10 in stock; Price: 1.5$
        • Dragonfruit:
          Availability: Out of stock; Price: 2$
  • Use Complete sentences: Complete sentences improve clarity. Here are two examples of an FAQ document:
    • Recommended:
      • Question: Can customers return the item within 30 days of receiving it?
        Answer: Yes, the customers can return the item within 30 days of receiving it, as long as the price tag is in place.
    • Not recommended:
      • Question: Can customers return the item?
        Answer: Yes.

You can use your existing Knowledge Sources to educate and optimize your AI Agents by following the recommendations and best practices mentioned in the article.

Feeling stuck? Do not fear.

Need assistance? Contact us for support. Help is here!

How to Write an Optimal AI Persona (Best Practices)

How to Align AI Conversations with Your Brand Voice

Can't find what you're looking for? 🔎

Talk to an expert here!

Contact Support
Company
About Us Roadmap Partner Program
Terms & Policies
Privacy Policy Cookie Policy Terms of Service
Resources
Developer Hub WhatsApp Pricing Calculator WhatsApp Link Generator
© 2024 Respond.io.
Status Security
ISO 27001 Certified.
Powered by HelpDocs (opens in a new tab)

Powered by HelpDocs (opens in a new tab)