Now that the product was in practical use by employees, we began to get feedback on real-world needs. The earliest versions had very limited search functionality consisting of a not-very-good keyword search and a faceted filter. This would work for more technical users or those with simple information retrieval needs, but it certainly wasn't going to adapt well to a growing user base and increasingly heterogeneous content.
To call up specific data, especially for visualization, the early versions required the use of SQL. This would have limited the potential users, which would have been the kiss of death for a product that was not even fully integrated into company workflows and was still seen as a trial.
We experimented with a natural language query (NLQ) search, however it ended up being too much complexity for an already complex project. And it was fortunate that we did not bite this off, because an off-the-shelf instance of ChatGPT would ultimately render something like this completely redundant.
Settled on imitation chat-style search
Instead, we implemented a very simplified abstraction layer on top of SQL that I had developed when working for Amperity the previous year. I called it "NLQ training wheels" because it superficially resembled natural language search despite not having the same level of flexibility. But the intuitive nomenclature of the commands made it far more accessible to non-technical users and created technological and cultural readiness for the true AI query engines that would come out in the years that followed.
Used social media idioms to facilitate learning
I designed three tiers of query complexity, and they included special characters based on commonly used symbols in social media such as @ signs and hashtags. The dropdown interface enabled experiential learning while emulating common social media search systems.
Delivered the new search feature
The next major release was based on ready access to any data with only a few keystrokes, and also included a focus on hotkeys and other expert-friendly features that matched the user needs we uncovered.