Combining Positions and Generations in Automatic Query Completion with Retrieval-Augmented Generation and Multi-Objective Alignment
Query Auto-Completion (QAC) is a key feature of modern search engines that improves search efficiency by suggesting completion as a user type. However, existing methods face significant challenges: traditional pipelines for recovery and positioning do not lend themselves well to the long tail and require extensive feature engineering, while the latest production methods suffer from optical illusions and safety risks. We present a unified framework that reformulates QAC as end-to-end list generation with Retrieval-Augmented Generation (RAG) and multi-objective Direct Preference Optimization (DPO).
Our approach includes three key innovations:
- Refurbishing QAC as a multi-purpose end-to-end inventory generator;
- A comprehensive approach including a RAG, a multi-purpose DPO with educated and rules-based verifiers, and an iterative review of high-quality transactional data processing;
- A hybrid service architecture that enables efficient production deployment under tight latency constraints.
Tests on a large commercial search engine show significant improvements: offline metrics show benefits in all dimensions, human testing reveals +0.40 to +0.69 popularity points, and controlled online testing achieves a 5.44% reduction in lockouts and a 3.46% increase in proposal acceptance, confirming that RAG is consistent and provides an effective solution for multi-level measurement. QAC.
This work represents a paradigm shift in generation-to-conclusion powered by large-scale linguistic models, RAG, and multi-objective alignment, establishing a proven production framework that can benefit the wider search and recommendation industry.
- † University of California, Berkeley
