Zepto's published search architecture only applies to logged-in users (medium.com)

Zepto has published a comprehensive engineering breakdown of their search architecture, titled “Building Search for a 10-Minute World,” showcasing the use of advanced technologies like Llama 3 for query correction and a mixture of experts for ranking. However, it was discovered that their search functionality fails for logged-out users, revealing a critical shortcoming in how their search system interacts with new visitors. Queries that returned blank pages for unregistered users performed flawlessly once the user was logged in, suggesting that their architecture may not be consistently applied across different user contexts. This issue is significant for the AI/ML community as it highlights the disparity between theoretical architecture and actual user experience. The flaws in Zepto's guest search experience—lack of fuzzy matching, absence of zero-results recovery strategies, and a malfunctioning fallback semantic search—illustrate common pitfalls in deploying advanced machine learning models in real-world applications. This discrepancy could severely impact customer acquisition, as potential users encountering poor search results may choose to abandon the platform. The analysis underscores the importance of testing systems in varied user states to ensure a seamless experience, reinforcing a common lesson for engineers: what is designed may not always reflect what is delivered.