SAP for Me: Generative AI Solutioning for Support Knowledge Search

[michelle_lewis-miller]

With more than 10+ million knowledge assets across the SAP ecosystem, it can sometimes be difficult to find the information you need. Think: multiple needles, multiple haystacks, in a hurricane. Over the last few years, we've been on a mission to change things. We've put in place a variety of intelligent search features in the frontend, including our support knowledge search, and many more in the backend. These solutions have dramatically improved the ease of finding just the right information.

But gen AI technology gives us an opportunity to do more and do it even better.

Enter our new generative AI search solutioning feature. This latest innovation on SAP for Me finds the most relevant answers to your query from across the SAP knowledge ecosystem—from help documentation to knowledge base articles—and compiles them into a single, concise solution at the top of your knowledge search results.

What is Different?

Our traditional intelligent search uses the keywords in your query, alongside a wide variety of sophisticated techniques and contextual data, to look across SAP knowledge assets to return a list of relevant results.  

With the new generative AI search solutioning enabled, you still benefit from that intelligent search functionality while also gaining the following:

• Improved traditional search result relevancy.
• A single, generated solution incorporating information from many different sources at once.
• A search experience optimized to work with natural language, not just keywords.
• Peace of mind that a solution generates only when the feature is confident it has understood your query fully and has something useful to reply.

We are also providing you with a few useful functions for feedback and engagement with the feature:

• Thumbs Up
• Thumbs Down
• Copy
• Toggle On/Off

Improve Your Solutioning Results

To get the best possible results from generative AI search solutioning, use natural language and context to make your query specific and clear.

For example, an ideal query might be, “Why did I get a SOAP class error message when creating an RFQ?” This query not only uses natural language but also provides context about what happened, when, and where. The answer that generates is very useful:

In comparison, the search “SOAP class error” may not generate a solution at all. If it does, the reply might not be specific enough to answer the query fully without additional refinement or research:

Intelligent search and knowledge teams are always learning. If no answer generated for you fifteen minutes ago, it might just generate now! Knowledge emerges and evolves constantly across the SAP ecosystem, and on average, we log thousands of interactions with that knowledge during every minute of every day. This mountain of data changes how the search behaves from one moment to the next.

When Will This Feature Be Live?

Soon! We are just wrapping up our final testing and plan to deploy the feature to SAP for Me in late Q1 or early Q2.

[ErikEbertDK]

Hi @michelle_lewis-miller … Q1 is gone so this will be in Q2. Do you know when ?

It looks really promising and I look forward to working with this as soon as possible.

Thank,

Erik Ebert - Pentos

[michelle_lewis-miller]

@ErikEbertDK :: We are now targeting a May 15th soft release for about 25% of search sessions, and May 27th for a full release to all search sessions (assuming the soft launch goes smoothly)! The solution is ready to put live right now (and the internal UAT results are very much tempting us to do so), but we want to do a few final optimizations on the backend and put a little UX polish on the design for ya first.

[ArjenVHooydonk]

@michelle_lewis-miller will it also filter out older KBAs which are probably not relevant anymore (now the most tedious part of working with searching for knowlegde base articles is trying to figure out if something in a release impacts the KBA)

[michelle_lewis-miller]

@ArjenVHooydonk :: No and yes. 😄 

We aren't hard filtering, but we are attempting to solve this challenge in a few other, more dynamic ways:

• Our query pipelines and semantic modeling is set up in such a way as to optimize for recency.
• Additionally, for those older documents that do pop up, the relevancy engine on its own will generally bury these if it determines that aren't entirely useful for the query in question.
• That engine is learning with each action in the search, as well, so if a user performs a search, sees an old KBA, and doesn't engage with it--it remembers that for next time and takes it into account, pushing down that item in relevancy.
• Finally, we have a few hundred (arguably thousands) of people responsible for these KBA, and they try to archive them once they are truly irrelevant to all customers. However, I want to acknowledge that we don't have this nailed at scale yet, as you've noticed!

Irrelevancy is a tough thing to monitor across so many varied assets. Our teams are incubating some very clever ideas internally to solve for this, however, and I can hardly wait to test them productively. My hope is that these will make a very marked difference to your experience in this area!