ASO Patents

Document type	Research paper, published by Apple
Title	Unifying Ranking and Generation in Query Auto-Completion via Retrieval-Augmented Generation and Multi-Objective Alignment
Authors	Kai Yuan, Anthony Zheng, Jia Hu, Divyanshu Sheth, Hemanth Velaga, Kylee Kim, Matteo Guarrera, Besim Avci, Xuetao Yin, Rajyashree Mukherjee, Sean Suchter
Affiliation	Apple, with one author at UC Berkeley
Venue	Preprint, 2026. arXiv:2602.01023
Source	arxiv.org/abs/2602.01023

What This Paper Covers and Why It Matters for ASO

The paper describes query auto-completion for a “mobile application search system” and is evaluated on a large scale commercial search platform. Apple does not name the App Store in the text, but two of the authors, Hemanth Velaga and Sean Suchter, are also authors on Apple’s App Store ranking paper, which cites this one.

What it describes is a replacement of the suggestion mechanism that US8364694B2 patented in 2007. That patent matched a typed prefix against words drawn from store assets and ordered the survivors by sales popularity. Retrieve, then rank.

In this system an LLM generates the entire suggestion list, in order, in a single pass. There is no separate ranking stage.

The stated reason for the change is the most interesting part for ASO. A retrieve and rank pipeline can only suggest what appears in historical logs, which makes it incapable of serving novel or long tail prefixes with sparse interaction history. A generative system has no such limit. The suggestion box can now propose a query nobody has typed before, provided the catalog supports it.

What the generator is given

Retrieval augmented generation means the model does not invent from nothing. Four sources are compiled into one prompt:

• Query Index: a fast lookup table derived from historical search logs
• Content Retriever: a search system combining textual and embedding based retrieval
• Catalog Metadata: titles, descriptions, ratings and popularity for each item
• Query Features: frequency and conversion rate for each query

Your title and description are in that prompt. So is your conversion performance on the query in question.

Six objectives, with explicit weights

The model is aligned with multi-objective Direct Preference Optimization against six verifiers, combined in a weighted reward the paper writes out in full:

• Relevance: do the suggestions cover plausible intents
• Safety: avoid unsafe or policy violating completions
• Engagement: the likelihood a suggestion leads to a downstream action
• Catalog groundedness: each suggestion must correspond to actual searchable content
• Context groundedness: suggestions must be supported by the retrieved context
• Diversity: cover multiple intents, avoid near duplicate suggestions

Engagement is defined concretely as a combination of the conditional conversion probability given the prefix and the general historical conversion rate. So conversion sits in the system twice: once as an input to the prompt, once as a term in the optimization target.

Diversity is the other term with direct consequences. Near duplicate suggestions are actively penalized, which means the list will not fill up with three variants of the same phrase. Competing for a concept is one slot, not several.

The numbers

Offline, for the large generator: coverage 93.0%, relevance 0.687, unsafe rate 0.65%, engagement win rate 6.58%, catalog ungrounded rate 0.49%, context ungrounded rate 9.43%, diversity 74.10. Human evaluation gave an item wise relevance score of 0.699 and a pairwise preference of +0.40.

The controlled online experiment is the headline, with all results statistically significant at p < 0.05:

• Characters typed fell 5.44%
• Suggestions taken rose 3.46%

A 5.44% reduction in typing across a store’s search traffic means a meaningful share of users are now completing a query the system proposed rather than the one they started writing.

What this changes for keyword research

• Suggestion presence is no longer purely a function of historical search volume. A generative system can surface a prefix completion with no log history behind it.
• Catalog groundedness means the suggestion has to resolve to real apps, and the model builds that resolution from catalog metadata. Your text is the raw material.
• Conversion rate is the only lever in the reward function that a developer can actually move.
• Personalization is not in the system yet. The paper lists user and session specific context encoding as future work, so suggestion research remains reproducible for now.

Paper Summary

Query auto-completion suggests completions as users type. The paper argues that traditional retrieve and rank pipelines have limited long tail coverage and need extensive feature engineering, while purely generative approaches suffer from hallucination and safety risks.

The proposed framework reformulates the task as end to end list generation using retrieval augmented generation and multi-objective Direct Preference Optimization. A suite of rule based, model based and LLM as judge verifiers scores candidate lists against relevance, safety, engagement, catalog groundedness, context groundedness and diversity, and those verifier outputs form a weighted reward used for alignment.

Training data is largely synthetic. A teacher LLM generates candidate lists, a critic LLM gives structured feedback on issues such as semantic redundancy, poor relevance and diversity trade offs, spelling inconsistencies and poor prefix matching, and the lists are revised iteratively. That pipeline produces roughly 50,000 prompt and suggestion pairs, augmented with human labeled examples amounting to about 10% of the training set.

A hybrid serving architecture allows deployment under strict latency constraints. Evaluation is stratified across head, torso and tail prefixes to reflect the production distribution.

A detailed breakdown of the verifiers, the reward formulation and the serving architecture will follow in a later update to this article.

Related reading: US8364694B2, the 2007 patent this system replaces, and WO2023235143A1 on the suggestions shown before the user types anything.