OpenAi+Setapp

Problem: Users faced challenges in efficiently locating desired apps, leading to frustration and decreased engagement. The existing app discovery process was time-consuming and lacked user-friendliness, impacting user satisfaction and platform retention. Solution: To streamline app discovery, we envisioned the Setapp Web Assistant, leveraging a conversational interface to guide users in finding apps based on features or use cases. The assistant would provide tailored recommendations, enhancing the user experience and promoting app engagement. My Role and Collaboration: As the Digital Product and Marketing Designer at Setapp, I spearheaded the development of the Setapp Web Assistant in collaboration with analysts, developers, and researchers. I facilitated communication between design and frontend teams, ensuring seamless integration of the assistant into the platform. Additionally, I sourced and processed data to train the assistant models, refining its capabilities through iterative testing. Approach: Our approach involved multiple iterations to fine-tune the Setapp Web Assistant's capabilities. Initially, we experimented with generating responses based on predefined prompts and app descriptions. Subsequent iterations incorporated advanced techniques, such as LangChain and embedding-based question answering, to enhance response accuracy and relevance. Results and Iterations: First Iteration: Initial attempts showed promising responses but required refinement due to hallucinations and formatting constraints. Second Iteration: Implementation of LangChain provided more coherent responses, yet token count inconsistencies raised concerns. Third Iteration: Leveraging embeddings and contextual chat information significantly improved response quality and relevance. Fourth Iteration: The final iteration focused on optimizing infrastructure to ensure efficient performance. Insights and Learnings: Chat-based models outperformed text-based ones, highlighting the importance of conversational interfaces. Techniques like "Putting word in GPT’s mouth" proved valuable for guiding user instructions effectively. Understanding and maintaining chat context improved user engagement and comprehension. Minimal infrastructure costs ensured cost-effectiveness in implementation.