The unsolved problem of enterprise autocomplete

This post details the technical challenges and solutions for building a low-latency enterprise autocomplete system. It addresses the 'cold start problem' for query suggestions by mining queries from permission-aware, translated document content rather than relying on query volume. For document suggestions, it uses a low-latency model to predict user visits and filters/ranks based on keywords, acknowledging the difficulty of optimal ranking when multiple documents share similar keywords. The system aims to reduce average time to satisfaction by 40% and currently drives ~20% of result clicks.