!Hero image describing Rozz search
Understanding User Intent: More than Just Words
A user’s query is a reflection of their intent, and understanding this intent is vital to providing a meaningful experience. A user could be looking for an overview of the website or seeking a specific form or link. The way Rozz responds should reflect this intent – summarizing content for broad queries, or pointing to specific links for precise requests.
One of the key technical challenges we’re tackling is teaching Rozz to accurately interpret user intent. Through sophisticated prompt engineering, we are improving how Rozz handles different types of queries. This involves adjusting parameters like the number of retrievals from our vector database, which contains embeddings of the website. By fine-tuning these parameters, we ensure Rozz can extract and present relevant information from the website efficiently.
Reflecting the Website’s Personality
A unique aspect of Rozz is its ability to adopt the “personality” of the website it’s used on, reflecting the tone and values of the company. This feature enhances the overall user experience, making Rozz feel like an integral part of the website. To achieve this, we’re working on the system message that includes instructions to help Rozz understand and the tone and style of the website, creating a cohesive user experience.
Presenting Results: Fast, Efficient, and Balanced
Rozz strives to present results in a manner that is fast, efficient, and easy to understand. We’re ensuring this by striking a balance between textual content and links. Rozz avoids the pitfall of being too succinct and merely redirecting users to the website, and also averts being too verbose and risk inaccurate extrapolations.
Achieving this balance requires us to control model parameters like temperature and max tokens. A lower temperature tends to generate too brief a response, while a higher temperature might yield overly creative and longer responses. Similarly, adjusting the max tokens affects the length of the response. It’s a delicate balancing act that we’re fine-tuning with every iteration.
Embracing Advanced AI Models
Rozz’s brain is powered by OpenAI’s GPT models. Transitioning from GPT-3.5 to GPT-4 has significantly improved Rozz’s performance, thanks to the enhanced comprehension and contextual understanding offered by the newer model.
Iterative Improvements: A Constant Process
We’re constantly working to enhance Rozz’s capabilities. Improvements in the retrievals, such as optimizing chunk size and overlap, are part of our ongoing work. We’re also developing Rozz’s ability to reference previous messages and follow up when needed, adding a level of interaction that most search tools lack.
Creating an effective tool involves a lot of trial and error, but it’s a process we’re committed to. The journey of building Rozz is marked by numerous iterations and constant learning, guided by our vision of creating a tool that truly serves the user’s needs and provides a seamless experience.
As we continue to refine and enhance Rozz, we’re excited about the potential it holds for revolutionizing the experience on websites, making the internet a more accessible and user-friendly place.
Posted July 5, 2023