Query-Aware Page Selection

Updated 9 July 2026

Query-aware page selection is a set of techniques that condition the selection, ranking, or retention of content units on the input query across diverse retrieval systems.
It employs multi-stage methods, such as coarse candidate filtering followed by query-specific refinement, to ensure efficient extraction of key information.
The approach is applied in web search, multimodal retrieval, prompt compression, and KV-cache optimization, yielding significant improvements in latency and relevance metrics.

Query-aware page selection denotes a family of mechanisms that condition the selection, ranking, or retention of page-like units on an input query. In classical web search, the selected units are webpage segments or snippet sentences; in multimodal retrieval they are document pages; in long-context LLM pipelines they are fixed-length prompt pages; and in systems work on autoregressive serving they are KV-cache pages. Across these settings, the central objective is stable: maximize query-relevant information under constraints imposed by latency, context length, storage, or downstream reasoning capacity (Kuppusamy et al., 2012, Wasserman et al., 28 May 2025, Hu et al., 20 Mar 2026, Tang et al., 2024).

1. Terminology and technical scope

The phrase has several distinct but related technical senses. In search interfaces, query-aware selection chooses the parts of a webpage that should be shown first. In retrieval and reranking, it chooses which document page should be promoted because it is most likely to answer the query. In prompt compression, it chooses which fixed-length pages from a long context should be preserved. In LLM serving, it chooses which memory pages in the KV cache should be loaded for the current attention query (Kuppusamy et al., 2012, Wasserman et al., 28 May 2025, Hu et al., 20 Mar 2026, Liu et al., 28 Aug 2025).

Setting	Unit selected	Primary objective
Web search presentation	Segment or snippet sentence	Show relevant portions of returned pages
Multimodal retrieval and RAG	Document page	Rank answer-bearing pages first
Long-context compression	Prompt page	Preserve relevant structure under token budget
LLM serving	KV-cache page	Reduce memory traffic during decode

A recurring property is that the query does not merely filter a corpus once. It actively shapes which units are treated as salient at the appropriate granularity. That granularity varies from sentence-level snippet extraction to page-level reranking to cache-block scheduling.

2. Search-era foundations: segment assembly and snippet extraction

An early explicit formulation appears in segmentation-based dynamic page construction, where the system does not present a linear result list but constructs a single dynamic page from segments extracted from multiple result pages (Kuppusamy et al., 2012). The pipeline is query-aware at two levels: only pages returned for the query are processed, and segments within those pages are weighted, personalized via a profile index $M = \{\eta_1,\eta_2,\ldots,\eta_u\}$ , and filtered by a threshold $\delta$ . The paper recommends VIPS, a vision-based segmentation method that uses both semantic and visual aspects of the webpage. Candidate segments are then assembled through template-based token replacement into one constructed page (Kuppusamy et al., 2012).

This formulation already contains several themes that remain central. First, relevance is finer than the page. Second, selection is conditioned not only on query terms but also on structural cues and, in this case, user profile information. Third, the selected units are optimized for presentation efficiency rather than only retrieval accuracy. The reported prototype results support this interface-oriented motivation: paired t-tests show significantly reduced time to fetch required information across all user levels, including Level I users with mean time 91.9 s for FEAST versus 64.1 s for the proposed model, with two-tailed $p = 0.001233$ (Kuppusamy et al., 2012).

A more modern search formulation is sentence-level snippet extraction. DeepQSE models snippet extraction as selecting the start sentence of a fixed-length snippet using a query encoder, a query-aware sentence encoder that jointly models title, query, and sentence, and a document-aware relevance encoder over the sequence $[g_q, g_s^1, \ldots, g_s^R]$ (Yi et al., 2022). Efficient-DeepQSE decomposes the task into a coarse-grained candidate sentence selection stage, where sentence representations can be cached, and a fine-grained relevance modeling stage based on a Cross Transformer (Yi et al., 2022). The two-stage design preserves effectiveness while substantially reducing cost: on English, DeepQSE is reported at about 1540G FLOPs and 31.9 ms, whereas Efficient-DeepQSE is about 132G FLOPs and 3.09 ms; their P@1 scores are 77.05 and 77.03, respectively (Yi et al., 2022). This establishes a durable design pattern for query-aware page selection: coarse cached selection followed by expensive query-conditioned refinement.

In multimodal RAG, query-aware page selection is stated directly as the reranker’s task: given a query, select and order pages so that the most relevant and answer-bearing page or pages come first (Wasserman et al., 28 May 2025). DocReRank addresses this problem by changing the training data construction. Rather than mining hard negative pages for each query, it generates hard negative queries for a fixed page, verifies answerability with VLMs, and trains a reranker to output “True” or “False” for query–page relevance, using $s(q,d) := p_{\text{True}(q,d)}$ as the relevance score (Wasserman et al., 28 May 2025). The empirical gains are explicit: on ViDoReV2 with ColQwen retrieval, baseline NDCG@5 is 58.6, MonoQwen reaches 64.4, and DocReRank-Full reaches 68.8; on Real-MM-RAG, baseline average NDCG@5 is 73.8, MonoQwen 80.9, and DocReRank-Full 86.3 (Wasserman et al., 28 May 2025). The effect is strongest where pages are visually or semantically similar but differ in fine-grained answerability.

Argus extends this idea from reranking to retrieval itself by making the page representation query-dependent. Standard late-interaction systems keep a fixed document representation $\mathbf{D}$ , but Argus introduces a region-aware Mixture-of-Experts so that the document representation becomes $\mathbf{D}(q)$ before MaxSim scoring (Abdallah et al., 3 Jun 2026). The query encoder produces both retrieval embeddings and a compact context vector; the page is pooled into spatial regions; and a router uses region content, position, and query context to select experts per region (Abdallah et al., 3 Jun 2026). This is still index-compatible late interaction, but the page representation is no longer query-agnostic. The reported retrieval numbers are correspondingly strong: Argus-9B reaches 92.67 NDCG@5 on ViDoRe V1 and 86.0 NDCG@5 on the combined V1+V2 leaderboard, and in an agentic retrieval pipeline on ViDoRe V3 improves NDCG@10 from 60.28 to 64.80 (Abdallah et al., 3 Jun 2026).

A structurally different formulation appears in ToPG, where page or passage selection is recast as navigation over a heterogeneous graph of propositions, entities, and passages (Delmas et al., 8 Jan 2026). Query-aware traversal is implemented with Suggestion–Selection cycles: a Suggestion phase performs query-aware traversal, and a Selection phase uses LLM feedback to prune irrelevant propositions and seed the next iteration (Delmas et al., 8 Jan 2026). In this setting, query-aware page selection is not simple ranking but iterative graph navigation toward passages connected by proposition-level evidence. The gains on multi-hop tasks are substantial: on MusiQue, EM/F1 improves from 19.5/30.3 in the naive proposition-retrieval mode to 34.0/47.0 in Local mode with three iterations; on HotPotQA, EM/F1 improves from 49.2/61.0 to 59.3/72.7 (Delmas et al., 8 Jan 2026). This shows that page selection can be fundamentally relational rather than purely pointwise.

4. Long-context selection and corpus-level query-focused analysis

In long-context prompting, query-aware page selection becomes a compression problem. BEAVER explicitly shifts compression from linear token removal to structure-aware hierarchical selection (Hu et al., 20 Mar 2026). A long context is segmented into logical segments, packed into fixed-length pages, encoded with dual-path pooling, and scored by a QueryPlanner that combines semantic and lexical branches,

$s_{\mathrm{mix}}(i) = \lambda s_{\mathrm{sem}}(i) + (1-\lambda)s_{\mathrm{lex}}(i),$

together with structural priors labeled Anchor, Flow, and Flash (Hu et al., 20 Mar 2026). Sentence-level smoothing then restores discourse integrity after page selection. The efficiency claim is concrete: on 128k-token contexts, BEAVER reduces compression runtime from 31.7 s for LongLLMLingua to 1.20 s, a 26.4× speedup (Hu et al., 20 Mar 2026). The relevance fidelity claim is equally concrete on RULER: at 16k context and a 3k-token budget, BEAVER averages 83.7, versus 47.9 for LLMLingua-2 and 28.8 for LongLLMLingua (Hu et al., 20 Mar 2026).

At the corpus level, the effect of selection strategy on downstream analysis has been evaluated directly. A systematic study over seven selection methods, four text analysis methods, two datasets, and 26 open-ended queries concludes that semantic or hybrid retrieval are strong go-to approaches that avoid the pitfalls of weaker selection strategies and the unnecessary compute overhead of more complicated ones (Rangreji et al., 13 Apr 2026). The hybrid “Direct Retrieval” strategy combines min-max normalized BM25 and SBERT scores by simple addition, while more elaborate procedures such as query expansion or MMR-based diversity reranking provide only limited or inconsistent benefits in this setting (Rangreji et al., 13 Apr 2026). The main implication is methodological: data selection is not merely a preprocessing convenience but a substantive determinant of which topics become visible. For query-aware page selection in exploratory or analytical systems, this argues for treating selection policy as part of the model, not outside it.

A common misconception in long-context systems is that processing more units is automatically better. The long-context compression results and the corpus-analysis study jointly contradict that assumption: additional pages can increase cost, dilute relevance, and alter the semantic distribution presented to the downstream model (Hu et al., 20 Mar 2026, Rangreji et al., 13 Apr 2026).

5. Query type, adaptive allocation, and representation-level analogues

A closely related line of work studies adaptive selection conditioned on query type. DIG separates queries into global and localized categories, showing that uniform sampling is both effective and efficient for global queries, whereas localized queries require query-aware selection (Li et al., 3 Dec 2025). Its training-free pipeline uses query-type identification, content-adaptive frame selection, LMM-based reward assignment, and refinement around rewarded segments (Li et al., 3 Dec 2025). Although the units are video frames rather than document pages, the result is directly informative for page selection: more selected units can harm performance when the query is localized. With Qwen2.5-VL-7B at 32 frames, DIG raises MLVU accuracy from 59.52 to 67.20 and LongVideoBench from 56.92 to 60.43 (Li et al., 3 Dec 2025). This suggests an analogous document strategy in which summary-like or thematic queries receive broad, coverage-oriented page sampling, while entity-, fact-, or section-specific queries trigger targeted page retrieval and local expansion.

Another relevant abstraction appears in query-aware adaptive dimension selection for dense retrieval. Here the selected units are embedding dimensions rather than pages, but the core mechanism is the same: construct oracle importance distributions from relevance labels, then learn a predictor that maps query embeddings to per-unit importance scores (Wu et al., 3 Feb 2026). The paper reports that the learned selector improves retrieval effectiveness over full-dimensional baselines, PRF-based masking, and supervised adapter baselines, with a fixed 30% retained-dimension setting serving as a robust default across datasets and retrievers (Wu et al., 3 Feb 2026). A plausible implication is that page selection can be formalized in the same way: pages become the discrete units over which a query-conditioned importance distribution is predicted.

These studies reinforce a broader principle: query-aware selection need not be uniform across tasks. It can be conditioned on query category, on the anticipated locality of evidence, or on the subset of representational features most likely to matter.

6. Systems-level reinterpretation: selecting KV-cache pages

In LLM inference systems, query-aware page selection takes a different but technically precise meaning. Quest treats the KV cache as fixed-size pages and estimates the criticality of each page from the current query vector by maintaining per-page minima and maxima of key coordinates (Tang et al., 2024). For a page $p$ , the score is

$S^{(p)} = \sum_{i=1}^{d} \max(Q_i m_i^{(p)}, Q_i M_i^{(p)}),$

an upper bound on the largest possible attention logit within that page (Tang et al., 2024). The system then loads only the Top- $\delta$ 0 critical pages for attention. At 32K context and token budget 2048, Quest reports a 7.03× self-attention latency reduction versus dense FlashInfer; end-to-end decoding improves by 1.74× in FP16 and 2.23× with 4-bit quantized weights (Tang et al., 2024).

TinyServe develops a closely related formulation using bounding-box metadata for each KV page and a fused CUDA kernel that combines page scoring, sparse memory access, and masked attention in one pass (Liu et al., 28 Aug 2025). Here too, the “query” is the attention query vector, and the “pages” are cache blocks rather than document pages. The reported gains are up to 3.4× speedup and over 2× memory savings with negligible accuracy drop (Liu et al., 28 Aug 2025). This systems interpretation clarifies that query-aware page selection is not restricted to semantic IR. It can also denote any query-conditioned decision over paged representations where loading all pages is unnecessary or prohibitively expensive.

Across the literature, several limitations recur. Verification remains imperfect in synthetic hard-negative generation for rerankers (Wasserman et al., 28 May 2025). Query-type classification and selection windows can fail on localized tasks or over-expand context (Li et al., 3 Dec 2025). Structure-aware compression still faces granularity and hyperparameter trade-offs (Hu et al., 20 Mar 2026). Query-conditioned retrievers such as Argus inherit storage and latency costs from late interaction and add interpretability challenges at the expert-routing level (Abdallah et al., 3 Jun 2026). Even so, the common trajectory is clear: query-aware page selection has evolved from interface-driven segment extraction into a general principle for allocating attention, storage, and reasoning capacity to the parts of a corpus or memory state that matter for a specific query.