Let's break down AI Search

CCB1115-K22

## Transcript X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:0 [MUSIC PLAYING] Hi, and welcome to the session called let's break down AI Search and how it works behind the scenes. I want to talk to you about AI Search. and what it is, but I'm not going to walk you through how to set it up or even show you what it looks like. I am, though, going to walk you through some of the main mechanics on how AI Search does what it does, and by doing that, give you a better understanding of how you can shape your content to provide better value for your end users. So the agenda is this. So we're going to quickly just look at what AI Search is, then we're going to learn about indexes, something called word stemming, and search term relevancy-- how to make your search relevant. And then we're going to look into how AI Research deals with typos, and in the end, look at how this matters. Why do we need to know these things to work with AI Search? So allow me to first introduce myself. My name is Anders Figenschow. I am a certified master architect, and also a ServiceNow developer MVP. I work at Sopra Steria as CTO and digital workflows, and I have about seven years of ServiceNow experience. But I also have 22 years of experience as a software developer. And if you hear me talking, just like now, my experience of those two years are heavily influencing what I'm talking about. So one of my last projects now was to work with Inter IKEA with their transformation of document management. And if you're attending the Hague this year at Knowledge, please do come by and have a look at that session as well, which I'm having with that customer-- as well as I have an in-person version of this session, which is a bit longer, and you can come and have a talk with me after. So what is AI Search? So it is the next-geneation search engine from ServiceNow. The previous one was Zing Search, and it's-- this new one is being made from the bottom up, so it doesn't really re-use anything from the old search, and then has a lot of new capabilities that improves it a lot. We're going to look at how that works. To install AI Search, you have to specifically ask for it. But you can do that from within your instance. You write AI Search status in the search-- in your instance, and it will give you a place. You can just click and request for it to be installed, and within a day or two, you'll have it ready to use. And you can just set up the search sources, and you're ready to go. So let's have a look at the architecture of AI Search. So there's data first-- so for example, knowledge articles, catalog items, users, or whatever you want to be accessible for the search engine as a search source. So the index is in the middle, and there's a reason for. That is the core engine, is the index, and we'll have a look at that in a second. You can also index external sources. So if you want to index data from, for example, SharePoint. You can set that up as well. And it's using API to grab the content, place it into the index table. It does not copy the articles itself, and it will not display it within ServiceNow. It will actually redirect you to where that content is originally. Then you have the service portal and mobile apps, and so on that is utilizing the AI Search-- which is basically a front end for the search engine itself And of course, you have the technologies that are underlying technologies that helping you out-- the machine learning the natural language and so on. So the features of the search engine AI Search is these index tables. It has typo handling, word stemming. It has some algorithms to create the scoring relevancy for a search. And it's something we'll look into also. And there's two things I want to mention while we're on this slide, which is the stop words and the synonyms-- so the synonyms first. So if someone is searching for a broken laptop in the search portal, in the Service Portal. Maybe it doesn't say laptop in that article. It says computer-- so broken computer. So instead of having to index every single variety of this to make your end users find the article, you would add a synonym of laptop to computer, and the search engine understands this, and it will find that article with the same sort of relevancy. The stop words is the opposite. So you have a dictionary-- and it's already in your instance-- that is containing words that are never relevant for a search, because if you have a word that is in each and every article, it carries no weight. So if the word me is in all articles, then searching for help me doesn't really work in the search engine. So you have to stop words that are basically excluding these very common words, so it doesn't mess with the weighting, and the scoring system, and the relevancy. Of course, if you're into that sort of thing, you can use stop words dictionary to censor. So you can maybe censor articles that you don't want people to see-- which, if you're in a country with as popular, that's what you do. Let's have a look at the indexes and how that works. So the index is the core of the search engine. It is a collection of all of the words present in your content library-- for example, the knowledge base-- with the listed scoring to each one pointing to each article. Now, it works the same way in, for example, a cookbook. So if you have a cookbook, you want to open up a cookbook and find a specific dish you want to make. You don't browse each and every page of hundreds of pages potentially to find your recipe. You would look in the back. And here's one example of a table of contents. So maybe I want to make light and fluffy waffles. So it's on page 21. So if I switch each and every page and go to page 21, it takes a long time to find it. So instead, I can just browse this very quick index and find that, oh, what I'm looking for is on page 21. There you go. There's my target. So I use that. So an index table in ServiceNow and other search engines as well is working the very same way. So it's creating a list of all the words that are relevant for you and telling the search engine where to find them-- so which article has this and that word or that term. And it will just be very much quicker to search through a smaller table with less data. And having less data is actually very important. So you need to, when you're building your index-- and how ServiceNow does this as well-- is to try to reduce the need to have these words. So stop words is one way. You don't need to index something that is in the stop word list. But you also need to take into consideration that you have variations of words. So when we get to that, we have something called word stemming. So let's look at one example of variations of words where you don't want to index all of them. So the word consult, that is the stem word for these variations-- consultant, consulting, even the plural, consultants. So instead of taking all of these words and indexing them, making the index potentially eight times bigger than, you have to making this slower to search-- then ServiceNow will just index the stem word consult. And when you're searching for consulting, it knows the stem of that word also, and it finds just the stem word. So it treats them as if it's the same one. And it will list up articles that has all these variations, and that's very useful as well. I have some code, actually I want to show you-- how it works. So there was a guy called Porter, who created an algorithm called the Porter stemming algorithm. So I found this just by looking around the internet. And it was fairly easy to find people who already made JavaScript versions of it. So I put it into a Script Include, and I want to look how-- see how it works. So in this one I'm just having-- inserting a word and finding the stem of that word. That's everything this does. I'm not going to show the code, because we don't have much time. So we just have to look at the results. So I'm going to try the word computerize. What is the stem of that word? So we can see here that the stem of computerized is computer-- makes sense. Let's see if we get the same result if you take the plural of computer, computers. No, actually not. So it's compute. So there's some surprises when it comes to these algorithms, but there's reason for it. We can even try to do something else. We can take computerism, if that is a word. And the stem of that one is computer as well. So with these examples, you can see how it shortens the world and finds the stem of it. There's a few words that-- not possible to use an algorithm for. For example, if you have words like verbs, like I am, you are, those are transformative, so they would be a little bit different. So in those cases, you have the most common ones are actually being translated into one of the forms of that word. And that's how you can use the Porter stemming method-- and how ServiceNow does as well-- to find a stem of a word. So then there's the search term relevancy. So when you have these words in the index, ServiceNow and the search engine needs to know your search query's relevancy to the different articles. So if you write broken laptop, some articles might be more relevant than the others. So how do we do that? So there are some algorithms again. So there's one call to TF-IDF algorithm. And ServiceNow is using this, and several other search engines are using the same one. So what it does is it's calculating the frequency of a word within the document. Let's say you have a document with 100 words, but the word laptop is mentioned twice. So you have a frequency of two. And then you have the inverse document frequency, which is basically count the frequency of a word across the document. So you can see, how relevant is the word in one document compared to the relevancy across all the documents? And with this algorithm, you can actually see that it gives some pretty good results when it comes to these search queries and which one's the search engine promotes to be the top match. And it places this into the index with the word. So maybe the word laptop is a pairing three articles, but one is more relevant to the other, because it has a higher TF-IDF score. That is the relevancy score. So the problem with people is that they can't always type correctly. And me being Norwegian, English being a second language, I type wrong all the time. So for these search engines to have type handling, it's helping me out a lot. So how does this work really you cannot add all the variations of typos into the index. It would make the index too big. So you need some dynamic algorithm to deal with these typos. There's, luckily, some smart people already figured out how this works, and there's something called string metric, or edit distance. What that means is that you have two words and you compare them to each other of how far away they are. What that means is, how many changes do you need to make to one word-- changes to the characters in that word to match the other one? So in this example, we have the word intention here, and you have another word called execution. One change is you can delete one character here, you can substitute these three characters, and you can insert another one here, and those five changes will transform the first word into the second one. That will give this a distance score of five, according to one of these algorithms called the Levenshtein algorithm. It's not documented ServiceNow is using the Levenshtein algorithm, but I'm assuming they do, because if you look into the Script Includes, you find the mention of Levenshtein a couple places. So I went and found that algorithm and tested it out in my own instance. So how about we have a look now and see how that works. So first let me show you that. I'm just making an array, and I'm just adding a few words into that array, just as a proof of concept now. And I'm going to look at a few of these words, but I'm going to write them as typos and see how the Levenshtein algorithm or the string metric at a distance algorithm works and gives me the best results. So first let me try to type in computer, but I'm adding an extra r. Like I told you, you don't want to index this word, because it has a typo. So I'm going to look through my array of words and see what the distance is for the nearest word. So here it says, you were looking for a computerr-- with an extra r-- and the closest word in the dictionary is the computer. And it has a Levenshtein distance of 1, which makes sense, because I just needed to delete this one. Let's try another word. Let's write comutre. I'm not sure this is a word. But it says that the closest word is commute, and it has a Levenshtein distance of 2. So this is the closest word. It realizes that this is probably a typo. Now, of course, as you can see, you have a much farther distance for all these other words ServiceNow is the distance of mine. So it's very far away. A good way to think about this is that you've put a threshold on it, and everything above, for example, 2 or 3 is probably not a typo, but a different word. So if I write something gibberishy, like this, let's see what-- it will find the closest word, but it probably won't be considered a typo. So the closest word is-- actually, several ones have this same distance, but 8 it is most likely not a typo. It's a different word. So this is how ServiceNow also uses algorithms to deal with your end users' typos and still give a good result. Even Google has this. If you write something which has a typo, it will ask underneath, did you mean, and the word closest to that one. And it will actually give you the results without the typo itself. So what? Now that we know how the search engine deals with typos, how it scores for relevancy, how it stems words and uses that in the dictionary, and how the dictionary's built, then so what? Why does this matter to us? Do we need to know these things to use the search engine? Well, obviously not, but it will help you to deliver better value to your end users if you know how it works. Let me explain. You can now advise your customers on how the relevance scoring works. So if you have an article with 100 words, if you have 10 words of those being the same, it will score very high in the relevancy scoring. So you can now tell your customer or end user that the frequency of word matters when it comes to the relevancy scoring in the search engine. Now, you can also tell your customer and user how ServiceNow handles the word variations. You don't need to put all of these variations in a meta field, for example, like computer, computerized, computers. ServiceNow will deal with that. So you can advise them to spend the time working with a search in other ways. Experimenting with stop words can give you totally different amount of results, and also using synonyms will improve your hit rate, so-- using laptop, and computer, and so on. So both stop words and synonyms already have tables in ServiceNow where you can have a look at some demo data as well. So by saying that, let me conclude and say thank you for checking in. And I'm hoping to see you in the Hague this year. Have a chat with me, and let's talk about how search engines work and how to make your life better for you and your customer. Thank you. [MUSIC PLAYING]

Watch Recording

From setback to success: Navigating challenges in ServiceNow

CCT2884

<p>Join this powerful session with a ServiceNow MVP and Certified Master Architect (CMA), as he shares his story of multiple setbacks and how he didn't allow that to define his journey, but used it as a catalyst to grow. Dive into how this MVP rebuilt his professional confidence and transformed his process to learning and problem-solving within the ServiceNow ecosystem. This session explores how setting ambitious goals, even beyond your current confidence level, can lead to significant personal and professional development.</p>

Watch Recording

Let's create Machine Learning from scratch with JavaScript!

CCB1121-K23

<p>Have you ever wondered how Machine Learning (ML) actually works? In this session you will learn how to create a simple ML model from scratch in JavaScript, so it can be easier to understand how it actually works. Expect to learn about the core principles of Machine Learning, and to see some fun examples of usage.</p>

Watch Recording

Teaching Your Instance To Fix Itself With Agentic AI

CCB6161

Every admin wants an instance that doesn't just report problems but resolves them. In this session, we'll explore how Agentic AI can analyze logs, identify recurring issues, and even generate fixes before users notice something’s wrong. You’ll learn how to design self-healing logic on ServiceNow that transforms maintenance from a reactive chore into proactive system intelligence.

Inside the black box: Generative AI technical deep dive

CCB1289

Curious about how ServiceNow’s Generative AI features work behind the scenes? Join us for a concise, yet insightful, exploration of the neural networks and transformer models that power code generation, ticket summaries, and other AI-driven capabilities in the platform. We'll break down the technical concepts, showing how these foundational technologies optimize workflows in ServiceNow. Plus, we’ll explore how understanding the tech behind GenAI could shape its future role on the ServiceNow platform.