MantisGem

What it does

Our app remedies a common problem that most users of LLMs face. Many users are interested in getting recommendations for activities or exploring options that are suitable to their specific needs given their general preferences or particular circumstances at the moment when they are querying. However, they are often stuck with responses that are too generic to serve their specific needs.

A skilled human assistant would remedy this as follows: ask them pertinent questions that help them to narrow their options, and gather information that may not be directly relevant but could help understand the user's thought process. The former relates to flipped interactions and chain-of-thought prompting, which have been areas where AI has advanced significantly. The latter, on the other hand, relates to tools from cognitive psychology and behavioral analysis.

We have utilized the ability of Gemini to perform the former, where we take the user's queries and combine it with a request to perform flipped interactions. This is done through API calls. The latter is more delicate and not straightforward, and hence we have used several handpicked training examples designed using a combination of Gemini and ideas inspired by the psychology literature. Thus our model utilizes innate abilities of Gemini and questioning patterns suggested by the examples to direct the conversation with the user in a way that best extracts relevant information from the user and guides them to the desirable output.