Improve Your Knowledge Base for Retrieval Augmented Generation (RAG) With These 10 Tips

In the age of Generative AI, knowledge management is more essential than ever. By grounding your generative content on your organization’s trusted knowledge base, a process also known as retrieval-augmented generation (RAG), you ensure that your generative content is accurate and specific to your organization’s processes. But creating knowledge for AI consumption requires a mindset shift on how you author, structure, and manage your knowledge content.

Based on my work as a Product Manager for our knowledge-grounded Generative AI products, and my many conversations with customers and experts, I’ve created a list of ten tips you should follow when creating and maintaining a knowledge base that’s used for Generative AI grounding. While this list is by no means exhaustive, it can help you get started on your Generative AI journey and deliver the awesome AI-powered customer service experiences that your customers have come to expect in 2024!

#1: Explain Thoroughly

Let’s be honest: we humans have short attention spans. When writing for mere mortals like us, it’s best? to be brief, to-the-point, and use simple language. On the other hand, Generative AI works best when given more thorough and complete information that it can then synthesize and alter for the appropriate audience. For example, customers who are chatting with a bot often want a brief and simple answer, while an agent who’s diagnosing a complicated issue may want the AI to get technical and detailed. That’s why I recommend being more thorough and detailed when writing knowledge content that will be used for Generative AI grounding. Err on the side of providing detail, and don’t be afraid to get technical. If you really want to be a Grade A student, I also recommend explaining common synonyms and abbreviations in your article. This helps the LLM to better understand how the different concepts in your knowledge relate to each other.

#2: Provide Examples

As you’re writing thorough and detailed content, you get bonus points for giving examples. If you provide scenarios that are conversational in nature, AI can interpret those when they’re relevant to the user’s query. When you’re writing a knowledge article about an issue, think of the situations in which a typical user would face that issue, and describe them in your content. This has the added benefit of putting yourself in your customer’s shoes, which is guaranteed to make your writing better!

#3: Structure Your Content

If you interpreted the first guideline as a license to create a word salad, get a fork. Just like us humans, AI likes structure. This means your sentences should be logically related to each other, and you’ll want to break up your content into paragraphs or lists when it makes sense to do so. I also recommend using heading tags (H1 to H6) to indicate how your content is hierarchically related.

But structuring goes beyond the way you logically relate your sentences. It starts at a foundational level with how you structure your knowledge content. In Salesforce, you can use record types, page layouts, and fields to establish a structure for your knowledge articles, and distinguish between different types of knowledge content. An FAQ article might be structured into “Question” and “Answer” fields, whereas a troubleshooting article might have fields like “Issue”, “Environment” and “Resolution”. Breaking your content down into fields and naming those fields appropriately doesn’t just help the writer by providing a template, but also helps the AI interpret the content you’ve written. I also recommend keeping your internal information and customer-facing information in different fields. This allows you to configure Einstein such that Reply Recommendations will only be grounded in customer-facing information, whereas your service agents can access information about internal service procedures when they’re chatting with Copilot. Showing the right information to the right audience is essential for maintaining trust and creating effective AI-generated content!

#4: Annotate Your Media

You were probably taught in school that an image is worth a thousand words. While that’s true for humans (most of us absorb information better and more quickly through images and video), the reverse is true for LLMs! It’s great to have screenshots, videos, and animated gifs in your knowledge content, especially for customers or agents who want to dive into a cited article for more information. If that content will also be used for GenAI grounding, I recommend annotating your multimedia with textual descriptions. Use text to explain what’s shown in the image or video, and leverage alt tags for annotations. That’s not just good for AI, it also makes your content more accessible for those who use screen readers. That’s what I call two birds with one (AI-generated) stone!

#5: Explicitly Address Your Customer’s Common Questions

This one may seem obvious, but hear me out. If you think about how users interact with Generative AI, you realize that most prompts come in the form of questions. The easier you can make it for the model to link those questions to their respective answers, the better AI will be able to help your agents and customers. Writing FAQs that explicitly address specific questions and issues greatly benefits both generative and non-generative use-cases! For example, Service Replies for Messaging uses the last 10-20 chats to understand which knowledge content to ground on. If your article content explicitly covers what a customer could have asked in each chat, you’ll help the AI find the right information more efficiently!

When you’re starting your knowledge base from scratch, my primary advice would be to start small and focus on what questions your customers are asking most often. I’ve spoken to hundreds of service leaders over the past few years, and many of them see creating a knowledge base as a daunting, insurmountable task. The reality however is that it doesn’t have to be hard! If you don’t have detailed analytics, you can still assemble a small group of your best service agents, and ask them to come up with your customers’ top 10 most common questions. Even if you just write ten articles to answer those questions, chances are you’ll already be able to solve a significant part of your incoming cases using Generative AI! You can then keep expanding your knowledge base over time, for example by leveraging KCS principles (discussed below) or using our Einstein Knowledge Creation product.

#6: Create Scoped Content

This one might come as a surprise. When I tell customers to create scoped content, I’m referring to the best practice of creating articles that cover one single topic or issue, rather than jamming ten different problems together in one article. Since LLMs are uniquely equipped to make sense or long documents, this recommendation might seem counterintuitive!

However, creating scoped content has two key advantages. First of all, it allows you to better manage security and visibility. Different topics might be appropriate to different audiences, and you might want to prevent certain users from seeing particular information, either because it’s irrelevant to them or for security purposes. By segmenting different issues and topics into different articles, you can use sharing rules to control who sees what, either directly or through grounded content. Second, users who are presented with a generative answer or response might want to click through to the cited source for a deeper dive. In that case, nobody likes being dropped into a 100-page document. Finally, defining content at a scoped level helps you better understand which content is most effective. We offer tools to measure which articles were used most for grounding, enabling your knowledge managers to understand which content is most effective, and which might need to be updated or archived. Knowing that one huge article was used often for grounding doesn’t tell you a lot; knowing that an article covering one specific issue was used frequently is much more informative and empowers you to grow an effective knowledge management strategy!

#7: Follow KCS Principles and Methodology

Have you ever heard of Knowledge-Centered Service, better-known as KCS? All the cool kids do it! KCS is a set of industry-standard best practices on knowledge management that help you leverage knowledge more efficiently. I consistently see that KCS helps our customers create a knowledge base that’s consistent, accurate, and represents the collective wisdom of your team of experts, all of which improve Knowledge grounding! To learn more about KCS, check out their website.

#8: Conduct Regular Audits of Your Knowledge Base

While Generative AI is great at adding meaning to your knowledge content, it doesn’t magically fix it. So when you’re grounding on a knowledge base with information that’s incorrect and out-of-date, chances are your generative content will similarly be… quirky. That’s why we recommend conducting regular audits of your knowledge base.

Fortunately, Salesforce offers tools to help you do just that. By using Salesforce Flow, you can build business-specific logic on top of our Next Review Date field, that will help you to run reports on articles that need to be analyzed. Outside of the regular review cycle, our Lightning Feedback managed package lets your users provide feedback on knowledge content directly, after which it gets assigned to a queue for review by a human writer. And finally, you can use Data Cloud to report on feedback that was submitted on your generative content to understand when it was incorrect, and when review of the underlying knowledge is needed. Especially with Generative AI, a robust knowledge management practice is more essential than ever, so this is the time to invest in that!

#9: Unify Content From Different Sources

In order to answer even more of your customers’ questions, you’ll probably want to incorporate information from sources across your organization. You might have PDFs with service procedures in Google Drive, product documentation in Confluence, and even relevant information on your company’s website. Thanks to the power of Unified Knowledge, you can easily ingest all of those knowledge sources into Salesforce, and enable them for Generative AI grounding. Because we only keep a copy or reference of your external content in Salesforce, you can keep control of it in your source system and maintain the integrity of your content. Your unified knowledge base can be exposed across every customer touchpoint, so you can leverage all of your organizational knowledge at every step of your customer’s service journey!

#10: Don’t Forget About Humans!

While improving your knowledge content for AI consumption is important, don’t forget your human readers. All our knowledge-grounded products display citations, linking to the knowledge that the generative output was grounded in. This adds a critical layer of trust and allows the consumer to dive deeper into the knowledge content if they want to learn more. Therefore, try to find a balance between writing content that’s consumable by AI but still understandable and useful to humans. At the end of the day, AI is here to make our lives better, not the other way around!

Start Your Knowledge-Powered AI Journey

I hope these tips make you feel more confident as you embark on your knowledge-powered AI journey. Venturing into this new world might seem daunting, but as I’ve laid out here, you can start small and already begin reaping great benefits. With Unified Knowledge, Service Replies, Search Answers, and Copilot, Salesforce offers you all of the tools that you need to quickly deliver great Generative AI-powered customer service experiences. And with Knowledge as the basis of your generative content, you add that layer of trust that’s now more essential than ever. So use these tips to your advantage, and start impressing your customers and agents with AI-fueled customer service today!