Generating Query Recommendations via LLMs (GQR and RA-GQR)
Forum topic · 小凯 · 2026-07-05
Summary
This paper introduces Generative Query Recommendation (GQR), a training-free approach that frames query recommendation as a generative task solved directly by a large language model (LLM) via carefully designed prompts. Traditional query recommendation systems depend on large document collections and query logs, which are expensive to collect and unavailable in cold-start scenarios. GQR removes these dependencies by prompting an off-the-shelf LLM with a single example, while its extension, Retriever-Augmented GQR (RA-GQR), dynamically composes prompts by retrieving similar queries from query logs. Experiments on the Robust04 and ClueWeb09B collections show state-of-the-art NDCG@10 and clarity scores against two commercial search engines and prior methods: GQR improves NDCG@10 by up to ~4% on average, RA-GQR by ~11% and ~6% respectively, and a blind user study found the generated queries most engaging at ~59% preference. Source: arxiv.org/abs/2405.19749
Generating Query Recommendations via LLMs (GQR and RA-GQR)
- Paper: Generating Query Recommendations via LLMs (arXiv, 2024-05-30)
- Authors: Andrea Bacciu, Enrico Palumbo, Andreas Damianou, Nicola Tonellotto, Fabrizio Silvestri
- Area: Search Assistance / query recommendation
Key points
- Problem: Query recommendation systems are ubiquitous in modern search engines, but they typically require a large collection of indexed documents and extensive query logs. Query logs are expensive to collect and maintain, and building recommendations involves complex, time-consuming cascading pipelines. In cold-start scenarios, query logs and user data are simply unavailable.
- Approach — GQR: The authors reframe query recommendation as a *generative task*. Generative Query Recommendation (GQR) uses an LLM as its foundation and requires no training or fine-tuning. A designed prompt lets the LLM understand the recommendation task, even with a single example, making the system simpler and ready-to-market.
- Extension — RA-GQR: Retriever-Augmented GQR exploits query logs when they *are* available by dynamically composing its prompt with similar queries retrieved from the logs.
Results
- Benchmarks: Robust04 and ClueWeb09B collections.
- GQR achieves state-of-the-art NDCG@10 and clarity score versus two commercial search engines and the previous state-of-the-art, improving NDCG@10 by up to ~4% on average.
- RA-GQR further improves NDCG@10 by ~11% on Robust04 and ~6% on ClueWeb09B over the best competitor.
- In a blind user study, the system won ~59% of user preferences, indicating the most engaging recommended queries.
Why it matters
- Offers a training-free, prompt-based alternative to cascading recommendation pipelines, viable even in cold-start settings without query logs.
- Reuses a pre-existing neural architecture, lowering deployment cost.
- Retrieval-augmented prompting (RA-GQR) shows query logs remain valuable when available, combining generative flexibility with behavioral signals.
Notes
- Quantitative figures above are taken from the paper's abstract; consult the PDF for full tables and statistical details.
Related reading
- Asking Clarification Questions to Handle Ambiguity in Open-Domain QA (arXiv:2305.13808)
- DiAL: Diversity-aware listwise ranking for query auto-complete (EMNLP)
- Enhancing Discoverability in Enterprise Conversational Systems with Predictive Query Suggestions (arXiv:2412.10933)
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178208960