— PROJECT

Gemini Conversational AI response framework

Gemini, while being useful, was extremely verbose, and always given to overexplaining when asked even simple questions. In addition to this, it also tended to add information that wasn’t relevant to the information being asked. 

In some cases, users need short, simple answers to their queries due to device and/or environmental considerations when Gemini’s response can’t be the primary focus of their attention. In certain scenarios, Gemini’s well-meaning verbosity could actually be a hazard, especially when used during activities such as driving or running. Additonally, overly verbose responses in audio-only scenarios could become annoying, leading to customer dissatisfaction. 

Presented below is evidence of Gemini’s verbosity. Note the overabundance of information.

Efficiently concise – Only the facts, please  

The challenge here was to create a framework that would train Gemini to tailor its responses to be helpful, concise, and efficient while being informative across different use cases. It would need to be able to respond to the user’s prompt to the best of its ability without going off the rails with its response length.  

I collaborated with a colleague to create a framework based on actual Gemini responses, along with tailored synthetic data. The ideal outputs lengths and the information they would contain covered the most common use cases, as well as important edge cases.

The actual framework and associated deliverables are covered by an NDA, but this visual demonstrates some of the kind of thinking that went into our work.