Multi-objective neural retrieval for query autocomplete

Query Autocomplete (QAC) systems predict the best query suggestions based on customer typed prefix and other contextual signals. Conventional techniques employ the Most Popular Completion (MPC) method, where query suggestions that are popular and begin with the prefix (prefix aware) are retrieved from a pre-computed index. To account for contextual signals like past search activity of the user in the session, QAC systems incorporate a re-ranking step on top of retrieved candidates. However, this is sub-optimal as the retrieved candidates do not necessarily contain any session relevant query suggestions. We propose an efficient way to retrieve session relevant, prefix aware and popular query suggestions at the same time. We present a vector transformation technique to combine different objective representations into one which is then used to search in a pre-computed vector index at inference time. We show that our method improves recall@100 over MPC and other baselines by 13% to 15% on one e-commerce dataset and AOL query logs without incurring significant latency.