Presented at Activate Customer Service 2021. With the proliferation of deep learning-based products like Alexa, Google Home, and Siri, your customers and internal users are accustomed to getting fast answers to their questions. Imagine having the same power to quickly answer your customer’s questions on your site, either through a chatbot or plain ole search box. The key is providing the correct answer in the first position, no matter how the question is asked. That last part is key, as most natural language-based questions can break typical lexical search engines.
This session will showcase Smart Answers, the deep learning, semantic vector based solution that powers chatbots and search boxes to return answers in the first position. No more scrolling or re-phrasing your questions. Business outcomes include less support cases for your support team and happier customers getting quick answers to their questions. Another benefit – internal support agents can leverage the same solution so they can be more effective in closing cases in a faster manner.
Brian Land, Lucidworks Director of Solution Engineering
Brian Land: Hi, everyone, welcome to this session today. I’m Brian Land, I’m a leader here at Lucidworks in the Solutions Engineering Group. And today, we’re gonna talk about leveraging semantic search to get answers in the first position in the customer service sector. So we’re going to have a short presentation, then we’ll show a couple of demos in how that works. But first, I thought I’d say a little bit about myself, and a little project I had during COVID.
I did one of those custom Sprinter vans you may have seen on social sites, but I learned a lot, I loved it. Our family loves to camp, we surf, we mountain bike, and things like that. Mom lives in Santa Fe, so we travel there too. But what I learned in this project was a lot of studying on blogs, and groups, and YouTube, and it was a lot, right? You start with installation, then you see those vertical stands to hang your wall on, plumbing, electrical, even natural propane gas with on-demand water heater, and then cook stove top.
Had to drill a lot of holes, five different holes in the van, which is really scary. There’s one for the window. There’s one for an electrical charging port, and a fan, and also an air-conditioner. It’s hot here in Texas. I’m in Austin, Texas, and it’s 100 degrees this week already, and so, this is really key. But this is a great brand, from the manufacturer. But what I found out, that it took about six to eight hours of just studying and problem-solving on different groups, and YouTube, and Facebook, and chats, and everything. A lot of manufacturers’ manuals just aren’t there. There isn’t a lot of it. They’re not gonna answer every single question.
I do have a lot of electrical experience in my background, but I found out through those six to eight hours in studying that, the AC unit takes 110 volt, the remote takes 12 volt. You have to wire them separately to the remote unit. There is a RJ11 wire, and it looks like a plane, remember those old pop phones, plain old telephone service? That’s four wires, RJ11. Yeah well, that doesn’t work, I found out. It’s called reverse polarity. So the first two wires are flipped. And so, once I figured all that out in lots of different sites I was hitting, customer service has taken days to get back to me but they kept just referring me back to different areas anyway. But yeah, it was a lot of challenges, and it was not the greatest experience.
It’s all working now, and I’m really happy with the van but I know from what we’ve heard from talking to others, custumer support are really overloaded, right? They’re having millions of employees now working from home, even support teams, support teams are remote. So if you have a specific question, you can’t just walk over to a cubicle, here’s your subject matter expert, ask some quick question, and get back out. It’s really what we’ve seen is that it exposed a lot of weaknesses of the current tools and just created an explosion of online support requests, burdening support teams.
And so, that’s why we’re here today is how can we help? How can AI and machine learning help? Here’s a quote from Gartner: “Gartner believes by 2025, customer service orgs are going to embed AI in their multichannel engagement platform, and they’re going to get more efficient by 25%.” That’s a huge number, and I could see that. I see what’s available today on the technology landscape, and how customer service organizations are working today, and how machine learning can help. And we’re going to go through some demos on that.
So I wanted to pose a couple of questions. So what if your customers simply can ask their question and get the answer in the very first position, right? Not hunt through search results, and pages, and different blogs. And we think about, well, there are companies that do that today, right? We’ve been asking Siri for years a question, and we don’t get a long list of results, typically. I think I did early on, but most likely now, they’re getting it answered in the first position. Google Home, Alexa, same thing. There’s a lot of automation there. So how do they do it?
Well, one, the key technology is called dense vector semantic search. And so, we’re going to show a couple of demos of that within our product, Fusion. And so, let’s just talk about what it is for a second for those folks that don’t know. So it’s instead of a lexical search, or an inverted index, if there are search engineers out there listening to this, we’re not just collecting all the words and documents, and aggregating that. The dense vector, it’s keeping terms in this example, the three-dimensional graphic here—horse is close to an animal and also close to the dog. There’s locations, New York, Beijing, and Paris, I’m in Austin. I’m flying on the airplane next week, but other things. So you can ask questions, and so it’s purpose-built for those natural language questions in understanding and intent, much like a Google, right? You get answers as you’re typing in the Google search box; it’s dense vector but it’s also collecting signals. Who you are? What you’ve done in the past? Where you’re going? You turn on your car, you’re about to go somewhere, Google pops up like, “Oh, you’re probably going to work. There’s a 12-minute delay,” something like that.
So you can ask questions, how much does it rain in Seattle, right? How many horses are in the U.S.? There’s that context, there’s intent. And it’s just really great for this. In my example, how to hook up a dometic AC unit to a remote for a Sprinter van, right? I’ve been asking those prolonged types of questions and Google was okay, I could find it by stitching together six different blog sites, and it was fine, but companies can be a lot better than that, right? They own the domain, companies, manufacturers, distributors, retailers in any vector, right, or any market. They can deploy this dense vector semantic search. So they own the manual, they own FAQs, glossary terms, experts, and provide a better experience. And that’s the next piece of signals.
So merging dense vector semantic search with signals, right? You understand your customers. I was a customer of that air conditioning unit. It knows what I bought, it knows what I’ve searched on in the past, and also can bring in information from other systems like ERP, commerce systems, right? It should be able to answer my challenges a lot faster. What type of reverse polarity? What type of cables should work with this unit I bought? So that’s what we’re talking about today.
The product from Lucidworks is called Smart Answers. We’re going to show a demo here in a minute. So it’s an advanced, deep learning semantic add-on for not just search boxes, but also chatbots. So it’s not a chatbot, it can power chatbots. Chatbots are really great for the UI, and roles, and workflow, and things like “what’s the status of my order?”, and it can ping the commerce system, answer it. But it’s not good usually for answering long natural language questions like, “how do I hook up my dometic AC remote to my Sprinter van?”. So it’s going to be great for just providing contextual answers, using natural language, and we’re gonna show how to use it both from a customer experience on one side as well as an agent experience. It can be used for both, and that’s what this leads into this next slide.
Think about your digital customer service agent in the middle. They’re typically serving from many different channels, like web, email, chat, Q&A. And they may not even know what’s going on until a case gets created, and they see it pop up in their CRM system. And so on the right, we have that customer experience, when customers are searching, and browsing, they’re buying, they’re searching knowledge bases, maybe some recommendations as well. By the way, these are all capabilities of Fusion. They may hit on chatbots, and Q&A, and looking at the community content.
If they can’t find their answer, they’re gonna create a case. And so, those cases that do flow into the agent, the customer service rep, this is the agent experience on the left. They also have their own tools. They have their CRM, and they have knowledge bases, and possibly Q&A systems as well. So there are other sessions from Lucidworks that cover this entire experience or other pieces. But today, we’re gonna talk about how semantic search can help the chatbot, Q&A, and search as well as from a customer experience, as well as to the agent experience to increase their effectiveness.
So let’s go to the demo. So here is a demo of Databak. So Databak is a data recovery and backup software. Let me get my notes here. And we’ve ingested lots of different knowledge base articles, glossaries, FAQs. That’s what Fusion is really good at is indexing from multiple data sources and providing a unified experience. You can filter on the left of these facets, looking at forums, but let’s start asking some tougher questions that may break some search engines out there.
So let’s start with a chatbot. I’m gonna ask a question. This has integrated both search and the chatbot from Databak. This is by the way, Google’s Dialogflow chatbot, it’s not our chatbot but we do power it in this example. So let’s ask a question, “what happens if the private analytics upload fails?” So let’s go ahead.
-[Chatbot]: When the data upload fails, then the service will try to upload again after 60 minutes. Upload failures are reported in the DBD.log file.
Brian: All right, that’s one natural way to answer that or ask that question. I’m gonna do another way. I’m gonna do this voice to text, so I’m gonna turn on the microphone. [on microphone:] “The upload isn’t working for the analytics server.” All right, there’s the answer.
-[Chatbot]: When the of data upload fails, then the service will try to upload again after 60 minutes. Upload failures are reported in the DBD.log file.
Brian: All right, so the difference between the first question and the second question, I used failed versus isn’t working, it’s understanding that context and that intent. So let’s ask a third way. [on microphone:] “What should I do if uploading to the analytic server is unsuccessful?”
-[Chatbot]: When the data upload fails, then the service will try to upload again after 60 minutes. Upload failures are reported in the DBD.log file.
Brian: All right, so you see how the third way, it was much longer. What should I do if uploading to the analytics server is unsuccessful? Unsuccessful versus failure versus isn’t working, three different ways to ask the same question, but we got the right answer. So that’s semantic search. I’m going to ask another question.
This time, I’m going to use the search box. So “what’s a DataCell Group?” And DataCell Groups are a group of DataCell environments. And so, actually pulling from a glossary, this is actually another data source. You see the glossary on the right, we have forms, data source, looks like a knowledge base on the left. There are a few different data sources we’re pulling from.
But if we ask another question like “who can read data in a DataCell location?” Here we go. Any administrator in the organization who has a Databak Cloud account can see the dashboard and reports. Also notice that this is now pointing to an FAQ. So we’re also ingesting from an FAQ data source and answering that natural language question. So let’s go back to chatbot and go kind of back and forth. “What is a client computer group?”
-[Chatbot]: A logical grouping of client computers in which selected options can apply to all member clients.
Brian: All right, that’s good. “Why would I need to use a client computer group?”
-[Chatbot]: Once created, client computer groups can be used to perform the following tasks for all their members; create scheduled tasks, set user security, view job history, generate reports.
Brian: All right, great. So you see how that question is worded where it’s around intent why. Why would I need to use a client computer group? So that’s a difficult question that would break a lot of search engines, but because it’s semantic-based, it’s able to hit that on the first try.
Let’s go ahead and ask another question here at the search box. We’re kind of going back and forth but I’m going to get on looking for an expert, “who can tell me more about client computer groups?” There we go. Oh, it looks like there’s even another data source people. Alea Abed will likely know. She specializes in network client computer groups and virtual machines. And if you’ve attended any of our other sessions, we also have about 25 machine learning jobs. This is not a job, this is where we’re showcasing dense vector semantic search. But one of those machine learning jobs is to find the experts, subject matter experts, based on content creation, and authors, and things like that. So we’re pulling in there, if you’ve seen some of our other demos we have a lot of call-outs on the slide. If you’re looking for a specific document or subject, here’s some SMEs on the right hand side.
All right, so that’s our demo on Databak. We’ll quickly pivot, and we just have a couple more minutes, but here’s another demo on the internal side. So by the way, this Databak, this could be customer-facing. I forgot to mention that. You can ask questions, try to solve your problem as quickly as possible. If I can’t solve the problem, I’m going to create a case, and that’s going to flow into probably a CRM system, like Salesforce, or Zendesk, and things like that.
And this is an example of a portal for customer service agents. And so, we’re aggregating lots of different data from Salesforce, Confluence, JIRA, knowledge bases, there’s some analytics and things like that. We do have plug-ins for Salesforce and others, but I just want to show it more on this holistic environment. So I have an agent add-on, I’m trying to solve my problem as quickly as possible. And so, of course, you can search, you can facet. There’s some machine learning on most popular searches. My open JIRA tickets is doing recommended knowledge bases based on my open ticket but we’re not here to talk about that. That’s recommender job, that’s another one of our machine learning jobs out of the box.
But if I ask a question, I would demo here, so I can show Smart Answers, right? This one on the right here is that semantic dense vector answer. The one on the left is standard search, lexical search using an inverted index. So if I were to ask a question, something like “how do I improve relevancy with query rewrites?” On the left-hand side was standard search. Let’s see, Getting Started with Fusion Server is not really what I’m looking for. Improve relevancy is here, so that looks like a big boost for that. Use Predictive Merchandiser Search Rewrites, this is actually pretty close. And yeah, this is gonna have some of what I’m looking for. But on the right hand side, think about intent, how-to. And really query rewrite is a machine learning term, by the way, and so it’s Predictive Merchandiser is our tool to power these query rewrites. But I’m looking like how do I improve relevancy using machine learning in these query rewrites.
And so, the very first item that popped up, Enabling and Disabling Query Rewriting Strategies, and see how relevancy is affected in each individual strategy. It looks like a really good one. I’m trying to just try to get an understanding how machine learning is going to help as I’m looking to improve relevancy on my site. And of course, Predictive Merchandiser. That’s our tool, that’s another really good one that was also in here on the left. There’s even more for relevancy and in query rewrites as well.
So those are just a couple of quick demos. I know we’re kind of up to our time. I’m going to stop sharing, and maybe we’ll turn it over to the team here, so you can do some Q&A here at the end but it’s a pleasure walking through the presentation and demo. So I’ll just go ahead and turn it over, thank you.