Eh... Because keyword and BM25 do not work that well, especially for enterprise search? Many of us surely complained how bad enterprise search was, right? A com

Eh... Because keyword and BM25 do not work that well, especially for enterprise search? Many of us surely complained how bad enterprise search was, right? A company usually has a hodgepodge of documents. They don't link to each other that much so page rank does not work well. They have fewer than 1 click per day, if there is any at all. So signals based on user interactions will not be useful or will be expensive to build. We often search with keywords that show up in enough number of documents but not enough to make BM25 rank the matches high. Query understanding hardly exists in enterprise search, because building query autocompletion, query correction, query reformulation and etc all requires some level of semantic understanding. If it's not embedding based, it has to be stats-based, which requires enough volume of clicks and user evaluation that enterprise search can't afford. The list can go on and on. There is a reason that Google's search appliance never took off, and people are still complaining about quality of enterprise search. Amazingly, a true semantic search would work well. And to do semantic search in scale, you'll need a vector database.