SmartSearch Exact Matches

[jbeast]

Hi,

We've been doing some testing of Smart Search on our content. Generally, the results have been very positive, however, we wanted to feedback on one particular point.

Occasionally, we know the exact phrase we are looking for within a page. However, it doesn't seem possible using Smart Search to find it. For example, when searching using the phrase Clinical errors are usually found to have, the page that has this string of words is the 20th result.

It'd be great to have a Google-like search syntax where you could use quotation marks to indicate you want to find an exact phrase (although I understand this would be complex as you'd have to have a embeddings/keyword hybrid search engine). 

Thanks, Chris

[jperkins · Accepted solution]

Chris @jbeast,

We definitely hear you and know that this is desirable behavior. The current engine behind Smart Search is unfortunately limited in what it can do, but we have work in progress currently to replace the engine with a better one over the upcoming year. This work has already been in progress for more than the past 6 months and we hope to be able to start validating the new engine in Smart Search early 2026 via an Early Adopter Program. 

[AngelaJacobson1]

RIGHT

[jbeast]

Hey Jeremy @jperkins , thanks for the reply. Exciting to hear there are improvements in the pipeline. Chris

[pgo586]

Hi @jperkins ,  great to hear your plans for replacing the search engine. Similarly (but somewhat opposite) to what @jbeast is proposing, I've found  in my tests, that sometimes I want to *only* retrieve semantic matches and *not* keyword-based matches in my searches. Controlling the behavior of the the smart search API via parameters (that let you specify whether you want a semantic|keyword|both type of search) would certainly be ideal.  In our use case, we want to use this search API within a chatbot, and our tests find that the search performs pretty well when semantically relevant content is actually found. However, it performs badly when semantically relevant content does NOT exist, as it returns keyword-based matches that are sometimes completely irrelevant but are nonetheless returned with a high confidence score. So essentially what we need is to control what type of search gets done so as to avoid using the keyword-based matches in our use case. As @jperkins notes, you'd need to be using a hybrid search engine on your end to support this feature, but I believe that you already do to some extent, and if not, I certainly hope that your 2026 improvements include support for this ).  BTW, how do we become part of the early adopter program?