Unifying Ranking and Generation in Query Auto-Completion via Retrieval-Augmented Generation and Multi-Objective Alignment¶
Core public-data mechanism.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Apple. |
| First public release | 2026-02-01 (arXiv v1) |
| Author code | No official author code linked by the paper as checked on 2026-08-24. |
| Local adapter or method key | rag-qac |
| Local reproduction code | src/auto_research/reproductions/rag_qac/ |
Paper method and evidence¶
Retrieve completion evidence from a searchable catalog. SFT and DPO align relevance, safety, engagement, catalog/context grounding and diversity.
flowchart LR
A[Retrieve catalog evidence] --> B[Generate grounded completions]
B[Generate grounded completions] --> C[Align multiple objectives]Original paper figure¶
Figure shows evidence retrieval and multi-objective alignment. From the original paper; copyright remains with the authors.
Core equation¶
\[
R=\sum_k w_kR_k,\quad L=L_{SFT}+\lambda L_{DPO}
\]
Reported results¶
Section 7/Table 3 reports suggestions taken +3.46% on 10% production traffic; not comparable to local ranking.
Local reproduction¶
MovieLens-100K, 220 users/360 items, same split/catalog/seed 42: shared transition+content baseline versus independent mechanism. NDCG@10 0.0540→0.0516 (-4.48%), Hit@10 0.1091→0.1045. Validation selects only blend weight. Metrics. No private logs, production models, bucketing or serving. Public proxies cannot establish online lift.