Best practices to use your knowledge articles with Now Assist (generative AI)

[Ashley Snyder]

This article is aimed at providing some best practice content on how to structure knowledge bases and format article content so that it can be best consumed by Now Assist in AI Search and Now Assist in Virtual Agent. It is a living article that will be updated as the technologies and capabilities expand and change. This is not an extensive or prescriptive article on how to create templates or draft articles but provides guidelines on how to best structure and format content for consumption.

This article is Version 2.0 and will be updated with more leading practices as they are discovered and as changes in guidance are made as the NowLLM models are trained. For more on NowLLM, see this community article.

 

What is Now Assist?

ServiceNow® Now Assist uses generative AI that is designed to enhance user productivity and efficiency through conversation and proactive experiences. We recommend taking the Now Assist Essentials course to understand more about how generative AI can be used in the Now Platform. 

Large language models (LLMs) are complex machine-learning models that are trained on large datasets like websites and documentation to perform language-related tasks, such as text generation for case summaries and resolution notes.

 

Where can I use knowledge articles with Now Assist?

Currently you can use knowledge articles with Now Assist in AI Search, with Now Assist Q&A Genius Results to display concise, actionable answers from knowledge article results in Service Portal, Virtual Agent, Employee Center, and global searches.  You can also use Now Assist to Draft new articles, leveraging existing resolved Tasks to build out the content.  Now Assist can even be used to help you identify potential duplicates in your knowledge base, to help reduce confusing or conflicting information.  For more information see the product documentation.

It's important to note that all generative AI technologies may generate text summaries that may be inaccurate, omit key information, or include irrelevant or redundant text. All LLMs are known to have the risk of producing Hallucinations.

---

Architecture and governance best practices

1. Maintain a single source of truth and reduce duplicates

Reduce or remove duplicate articles. When using RAG in Now Assist for AI Search, duplicate articles that have different content may provide an incorrect or disjointed answer for the requestor. Maintaining duplicates in various knowledge bases can lead to outdated and inaccurate versions. To avoid this, try not to create duplicate articles across knowledge bases. Instead, control who sees articles by using user criteria at both the knowledge base and article levels. If you need to show articles across multiple knowledge bases, consider using Unified Taxonomy in Employee Center. This method displays articles across topics without duplicating content for each knowledge base.

For use cases where you need to show personas different information on an article, you can use field-level security on knowledge article fields. Field-level security can serve for use cases in which you need to limit content to specific audiences within a knowledge article. For example, if you have an article, in which different teams in your support organization (e.g. Tier 0, Tier 1, and Tier 2), should see different content- it is possible to separate that out in a single article. The obvious benefit is that you would not need to maintain multiple articles on the same subject matter for different audiences, which increases the chances of inconsistent information. We support this feature out of the box with article templates.

2. Using native knowledge bases in ServiceNow and 3rd Party Connectors with our Now Assist products

Currently, Now Assist in Search and Virtual Agent supports summarization of content from external sources. Knowledge Management provides the most benefit within our Applications when Articles within ServiceNow are used, because you are able to track views, usage and feedback records within the system.  However, AI Search with Now Assist also allows you to search and surface Results Lists and Genius Results from 3rd party applications, like Confluence and SharePoint.

3. Create hyper-focused knowledge articles and reduce links

Focus on one topic for each article, and cover it fully. The LLM can't click links or go to other pages to summarize info. Complete, relevant articles help the LLM summarize better for Now Assist in Virtual Agent and Search.

As an example, there are links to content within this community article. If this were to be summarized with generative AI currently, the pertinent content for generative AI to use would need to be added to this article in order to provide a robust solution to the user.

Clear language also assists in accessibility needs for those with cognitive or other disabilities, as well as those who do not have the article written in their first language. While LLMs can handle technical terms and jargon, think about the readers who are consuming the summarized articles.

4. Use concise plain language and avoid jargon, technical terms, and abbreviations

Write in clear and simple language that is easy to understand. Avoid unnecessary jargon or overly complex language that may confuse the generative AI model or detract from the clarity of the content.  Avoid abbreviations when possible, and write the term out in its entirety. You can include the abbreviation in parentheses when using it for the first time. Again, while LLMs can consume abbreviations, we want to ensure the user on the receiving end understands what the abbreviation means. 

5. For FAQs and Q&A provide definitive answers, context, and consistent formatting

Generative AI can parse FAQ articles to some extent, but there are challenges associated with these types of articles. FAQ and Q&A articles may pose challenges with generative AI understanding the context of the article. It's important to ensure the LLM has enough context to understand the content being summarized, which may include relevant background information, key concepts, and the purpose of the summary. When using FAQ and Q&A articles, ensure the article has enough context for the LLM.

For Question and Answer (Q&A) articles, ensure the answers are self-contained. For example, i.e. "Does Now Assist have additional licensing costs?" instead of answering the question as "Yes", answer the question as "Yes, Now Assist does have additional licensing costs."

Ensure your FAQs are structured and consistent. FAQs can vary in formatting, language, and content. Using article templates for FAQs will help to ensure consistency. Generative AI may struggle with unstructured data or unconventional formatting. 

6. Provide content in text alongside images, videos, and attachments

Generative AI models are typically designed to process and generate text-based content rather than audio or visual content. Therefore, formatting related to audiovisual elements such as timestamps, audio cues, or image descriptions may not be directly processed or preserved by these models. Make sure to write out content and steps in the text. Screenshots are great for people to look at when reading articles, but language models can't summarize what's in screenshots, or videos right now. So, it's important to have any instructions or processes shown in the text itself in addition to any images, or screenshots.

The NowLLM cannot summarize attachments currently, it can only retrieve the HTML content from the most relevant article on your instance to pass to the LLM as context. 

7. Ensure articles are relevant and up-to-date

Ensure that content is relevant and up-to-date. Keeping the content up-to-date ensures that the generative AI model generates accurate and relevant responses based on the most current information available. If you're using the Knowledge-Centered Service (KCS) methodology, your articles will be improved via the Solve loop. If you're not using the KCS methodology, here's some guidance on what to prioritize:

• High-traffic articles. Our out-of-the-box dashboards provide insight into article view and use counts, as well as the number of times viewed.
• Articles that have high helpfulness ratings. 
• Articles that rank high on other success metrics defined by your organization such as rank.

This is an area where your governance process is key in ensuring articles are relevant. There are no prescriptive criteria built into the product. Your governance process will determine when articles should be reviewed and the cadence of those reviews. The Valid To field on knowledge articles controls whether articles and be browsed, searched, and used by generative AI. Ensure the Valid To date of an article is set by your governance process and is not an arbitrary date.

There are out-of-the-box dashboards to assist in reporting on view and use count to assist in pinpointing higher trafficked articles for prioritization.

1. Knowledge Management Overview (No additional licensing)

2. Knowledge Management (Requires Platform Analytics licensing)

8. Maintain consistency and quality via governance

Maintain consistency in terminology, formatting, and writing style across knowledge articles in your knowledge bases. Maintaining consistency will help to ensure uniformity and coherence in generated content. Using article templates for your knowledge bases will help to maintain consistency when creating articles. If you're using Knowledge-Centered Service (KCS) the Content Standard will lend to this best practice in conjunction with article templates.

Your governance process ensures that the knowledge articles used by Now Assist are relevant, accurate, and reliable. Governance establishes the standards for data integrity and quality, which is essential for generating trustworthy outputs for your end-users and customers. Implement a quality assurance and review process to ensure the accuracy, relevance, and coherence of the knowledge articles. This may involve peer review, expert validation, or automated checks to identify and correct any errors or inconsistencies in the content.

Ownership groups can be associated with knowledge articles to maintain article quality, manage approvals, and for timely resolution of feedback, rather than relying on a single author to maintain article quality.

The Article quality index checklist can be used in either KCS coaching or as part of approval workflow processes if you're not using KCS to ensure content meets the knowledge base standard.

9. Utilize actionable feedback

Use actionable feedback and feedback functionalities on articles such as rating, flagging, marking helpful/not helpful for users to provide input on knowledge articles and suggest improvements or corrections. User feedback helps identify areas for enhancement and ensures that the knowledge articles remain valuable and effective for use with generative AI. For more information on feedback and actionable feedback, see our product docs.

---

Formatting best practices

1. Use a clear structure with titles, subtitles, and steps

Write knowledge articles in a clear and structured manner, with well-defined sections or headings that organize the information logically. This structure helps the generative AI model understand the content hierarchy and generate coherent responses. Utilize headings, subheadings, and bulleted lists to break down complex information into digestible chunks. This formatting makes it easier for the generative AI model to process and generate content based on specific sections or points within the article.

Titles

Phrase the title using terms for which customers are likely to search (the symptom) rather than the underlying technical problem. Try to be specific.

• For example, instead of "DNS Server is down" try saying "Cannot navigate to a website" For more information and examples, see the article Article Title Guidelines (KCS).

Abstracts

Create an abstract at the very beginning of the article, for longer articles, or articles that are structured in a more complex layout, to help AI understand what the main topic of the article is and leverage that abstract in Genius results.

Bullet points

Use bullet points or numbered lists to break down complex information into concise and digestible points. This format makes it easier for the AI to process and generate coherent summaries.

 

Links

Providing a detailed summary of the information within, or the intent of navigating to a link should be outlined in the article.  Tip:  If the links are structured in a style that outlines specifically what the link is for, and then the link, Genius results will often serve up the link in a given synthesized answer.  Example:  "For information on our host's parking pass policy, click: Parking Pass Policy"

2. Bold or italicize key terms

Highlight important terms or concepts by using bold or italic formatting. This draws attention to key information and improves the AI's ability to identify and summarize the main points.

3. There are formatting limitations

Some basic text formatting such as bold, italic, and underline can be processed by generative AI models, but more complex rich text formatting like fonts, colors, and advanced styling may not be recognized at this time. It's suggested to stick to plain text as much as possible in order to ensure a consistent interpretation of the article.

4. Maintain consistent formatting across articles

Maintain consistency in formatting throughout the text and throughout articles. Use the same font size, style, and spacing to ensure coherence and readability.

5. Use meta and tags on knowledge articles for relevancy

Now Assist in AI Search and Virtual Agent use AI Search technology. AI search utilizes meta and tags on knowledge articles for relevance (as of the San Diego release). For more information on AI Search and Knowledge Management best practices, see this session. Now Assist in AI Search will use the HTML fields on returned knowledge articles from AI Search.

6. Paragraph length

Keep paragraphs short and focused to maintain readability and coherence. Avoid lengthy paragraphs that can overwhelm the AI and result in fragmented or unclear summaries.

7. Use descriptive captions

Provide descriptive captions for visual elements to explain their relevance and context. This ensures that the AI can accurately interpret and summarize the information presented in the visuals.

8. Use text in conjunction with images and videos

Currently, generative AI only supports text. If you have content in images, see if the content can be added as text as well. This also supports users with accessibility needs, as users with visual disabilities may not be able to view images and videos. 

9. Minimize the use of table formatting

Generative AI may have issues with reading and preserving table formatting. While generative AI models can process structured text data to some extent, they may struggle with complex structured formats such as tables, charts, or databases. Converting structured data into a format that is compatible with the model's input requirements may be necessary for effective processing.

10. Non-standard characters have limitations

Generative AI models may struggle with processing text containing non-standard characters or symbols, particularly those outside the standard character set used in natural language text. This can include mathematical symbols, emojis, or special characters not commonly found in everyday language.

See our article on Style Guide Best Practices for more information.

---

FAQs

Are knowledge blocks supported in AI Search and Now Assist in Search/Virtual Agent?

A: Yes, knowledge blocks are now searchable by AI Search and can be used in Now Assist in Search/Virtual Agent.

What countries/languages are supported for Now Assist?

A: ​​​​​​​All currently supported countries on the Now Platform can use Now Assist. English, Spanish, French, German, and Japanese languages are supported through Now LLM. Other languages are supported through dynamic translation. The GAI Controller, which connects to Azure OpenAI, OpenAI, or Google Vertex AI LLM, may support other languages since those LLMs have multi-language support. Refer to their documentation for details. 

Are there technical limitations on the length of articles?

A: All LLMs have limitations on the amount of data that can be processed as an input, and the amount of data that can be generated. The inputs are usually a mix of instruction data (this tells the LLM what to do and what not to do) and context data (e.g. KB article retrieved by AI Search, or chat conversation to be summarized), and the outputs are generated data. Please refer to the respective model cards for discussions about other model-specific limitations.

I have more questions about Now Assist

A: See the Now Assist FAQs and Now Assist for Search FAQs

I have more questions about AI Search

A: See the AI Search FAQs and our AI Search and Knowledge Best practices session with associated Q&A

[Daichi Ishikawa]

@Ashley Snyder 

Thank you for a valuable information!

This article is so useful for me!!

I've been considered what a good knowledge article is for using generative AI. 

[Thom Schreck]

Is the articles published on date relevant to this process? For example, if an already well formatted article is 3 years old but still valid and useful, should we republish it to have a more recent last updated date?

[RussLaPlante]

Hello,

We're using Washington DC - do we still need to use meta tags in articles? Is this feature being depreciated with the improved searching of AI?

Thanks, Russ

[CMartin0423]

All these Best Practices are very helpful to read and understand as my team begins its Now Assist journey.

Does ServiceNow have any resources/suggestions we can explore, regarding how best to implement each of these Best Practices? We're primarily interested in learning how we can help our knowledge content owners make their existing articles more consumable by AI. What governance and automation can/should we put in place to accomplish this?

[MFuchsloch]

Hello. For point #7 in the Formatting Best Practices, to what extend will tables not be functional. Please see an example here of a holiday calendar in a table format. Is it better to include these details in an attached pdf or would this basic table format be readable by AI?

[GASHOK]

Thank you for sharing this! This was really clear and easy to follow. I especially liked the points on keeping articles focused and reducing duplicates so Now Assist can give better answers. We’re working on improving our KB for AI Search, so this helps a lot. 🙌

Looking forward to more best practices as NowLLM evolves. Thanks again for putting this together!