QChunker: Query-Aware Adaptive Chunking

• The paper introduces QChunker as a method that dynamically determines chunk boundaries using query-conditioned seed selection and contextual expansion.
• It leverages a dual approach: query-adaptive semantic chunking (QASC) and a multi-agent debate framework for segmentation and knowledge completion.
• Empirical results demonstrate significant improvements in Precision, Recall, and overall chunk quality compared to fixed segmentation methodologies.

QChunker denotes a class of question-aware chunking methods for retrieval-augmented generation (RAG) in which chunk formation is no longer treated as a static preprocessing step. In the 2026 arXiv literature, the label is attached to two distinct formulations. One is Query-Adaptive Semantic Chunking (QASC, also identified as “QASC or ‘QChunker’”), which dynamically constructs chunks by combining query–sentence similarity, contextual window expansion, chunk-level score aggregation, and boundary adjustment. The other is a domain-RAG framework named QChunker that reframes chunking as text segmentation plus knowledge completion under an “Understanding→Retrieval→Augmentation” paradigm, implemented through a multi-agent debate process and evaluated with a direct metric called ChunkScore (Rastogi, 29 Apr 2026, Zhao et al., 12 Mar 2026).

1. Conceptual scope and nomenclature

The two principal QChunker formulations share a common premise: chunk quality constrains the effectiveness upper bound of RAG systems, and chunking should be guided by semantic relevance rather than only by fixed-size heuristics. QASC argues that fixed chunking produces a precision-recall trade-off “unresolvable by tuning chunk size alone,” because segmentation is performed without integrating user intent. The multi-agent QChunker identifies related failure modes in specialized domains, including “term-definition gaps,” “missing background,” and “broken context dependencies,” and therefore models chunking as a task that must actively preserve logical coherence and information integrity (Rastogi, 29 Apr 2026, Zhao et al., 12 Mar 2026).

Variant	Core mechanism	Primary setting
QASC / “QChunker”	Query-conditioned seed selection, contextual windows, weighted scoring, merging	Technical-document RAG
QChunker via multi-agent debate	Question outline generation, segmentation, integrity review, knowledge completion	Domain RAG

These formulations differ in granularity and operating point. QASC is a query-time chunking strategy operating over sentence sequences. The multi-agent QChunker is a document-understanding pipeline that first generates candidate segmentations, then rewrites incomplete chunks so that they become self-contained. This suggests that the term “QChunker” names a broader research direction rather than a single standardized algorithm.

2. Query-adaptive semantic chunking

In QASC, a document $D$ is represented as a sequence of sentences $D=\{s_1,s_2,\dots,s_n\}$, and a sentence-embedding model $e:S\to\mathbb{R}^d$ maps both query and sentences into the same vector space. Query–sentence alignment is measured by cosine similarity,

$$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$

This produces a similarity profile $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$. Seed sentences are selected by an adaptive percentile threshold: $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$ In the primary experiments, $p=75$, so the top $25\%$ of sentences by similarity become seeds (Rastogi, 29 Apr 2026).

Around each seed sentence $s_r$, QASC constructs a contextual window

$$W_r=\{s_{\max(1,r-m)},\dots,s_r,\dots,s_{\min(n,r+m)}\}.$$

The default window radius is $D=\{s_1,s_2,\dots,s_n\}$0. The paper also describes an adaptive setup in which forward and backward expansion continues until similarity falls below $D=\{s_1,s_2,\dots,s_n\}$1. Candidate windows are then scored by a distance-weighted mean,

$D=\{s_1,s_2,\dots,s_n\}$2

with default $D=\{s_1,s_2,\dots,s_n\}$3. A chunk is retained only if its score exceeds $D=\{s_1,s_2,\dots,s_n\}$4, with $D=\{s_1,s_2,\dots,s_n\}$5. Overlapping or adjacent high-scoring windows whose spans are within a gap tolerance $D=\{s_1,s_2,\dots,s_n\}$6 are merged, and each merged chunk boundary is snapped by up to two sentences to the nearest paragraph or section break. The default embedding model in the main experiments is Sentence-BERT all-MiniLM-L6-v2 with $D=\{s_1,s_2,\dots,s_n\}$7 (Rastogi, 29 Apr 2026).

The significance of this formulation lies in where query information enters the pipeline. Rather than retrieving from a static inventory of chunks and only then filtering by query, QASC uses the query to determine the chunk boundaries themselves. The paper characterizes this as “query then chunk then retrieve,” and reports that the combination of seed selection, local context preservation, and chunk-level aggregation resolves the precision–recall trade-off that fixed chunking leaves unresolved (Rastogi, 29 Apr 2026).

3. Multi-agent debate, segmentation, and knowledge completion

The domain-RAG QChunker formalizes chunking as a composite mapping from document $D=\{s_1,s_2,\dots,s_n\}$8 to a final chunk set $D=\{s_1,s_2,\dots,s_n\}$9, decomposed into text segmentation and knowledge completion. After a candidate segmentation is selected, each chunk $e:S\to\mathbb{R}^d$0 is checked for omitted but necessary context. If the omission is material, a minimal set $e:S\to\mathbb{R}^d$1 is extracted and the chunk is rewritten as

$e:S\to\mathbb{R}^d$2

where $e:S\to\mathbb{R}^d$3 denotes a coherence-preserving rewrite rather than raw concatenation. The resulting output is $e:S\to\mathbb{R}^d$4 (Zhao et al., 12 Mar 2026).

This framework is implemented through four specialized agents. The question outline generator $e:S\to\mathbb{R}^d$5 interrogates the document and produces a structured outline of probing questions $e:S\to\mathbb{R}^d$6. The text segmenter $e:S\to\mathbb{R}^d$7 uses $e:S\to\mathbb{R}^d$8 for multi-path sampling, generating a candidate set $e:S\to\mathbb{R}^d$9, then selecting the optimum segmentation with ChunkScore. The integrity reviewer $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$0 compares each chunk against the full document and returns missing knowledge points $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$1 together with a flag $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$2 indicating whether the omission materially impairs comprehension. The knowledge completer $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$3 is invoked only when $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$4, verifies the missing points, and rewrites the chunk so that the final text is self-contained, coherent, and stylistically consistent (Zhao et al., 12 Mar 2026).

The pipeline operates sequentially as

$$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$5

and is explicitly presented as a shift from “Retrieval→Augmentation” to “Understanding→Retrieval→Augmentation.” Through this process, the paper reports construction of a high-quality dataset of 45K entries and transfer of the capability to small LLMs. The data are organized into three training corpora: $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$6 for outline generation and optimum segmentation, $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$7 for integrity discrimination, and $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$8 for chunk rewriting (Zhao et al., 12 Mar 2026).

4. Evaluation methodology and chunk-quality metrics

QASC is evaluated on 100 technical documents across 200 queries spanning four types: factoid, topical, comparative, and multi-hop. The RAG pipeline uses FAISS indexing of Sentence-BERT embeddings and GPT-5 as the generator. The main reported measures are Precision, Recall, F1, Answer Correctness, Faithfulness, and Latency. The paper also includes a blinded human study on 40 document–query pairs with three annotators, reporting Fleiss’ $$\mathrm{sim}(q,s_i)=\frac{e(q)\cdot e(s_i)}{\|e(q)\|\;\|e(s_i)\|}.$$9 of 0.78 for Relevance, 0.74 for Coherence, 0.80 for Completeness, and 0.76 for Answer Quality; pairwise Cohen’s $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$0 ranges from 0.72 to 0.84 (Rastogi, 29 Apr 2026).

The multi-agent QChunker introduces ChunkScore as a direct chunk-quality metric intended to avoid the long evaluation chains of downstream QA-based assessment. ChunkScore combines a micro-level measure of Logical Independence (LI) and a macro-level measure of Semantic Dispersion (SD). For adjacent chunks,

$$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$1

and the boundary-averaged score is

$$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$2

Semantic Dispersion is defined from chunk embeddings $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$3 through a centered Gram matrix $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$4, with

$$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$5

The final metric is

$$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$6

Empirically, $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$7 maximizes correlation with downstream ROUGE-L, and Fig. 3 reports a Pearson correlation of approximately 0.98 (Zhao et al., 12 Mar 2026).

For downstream task evaluation, the same paper uses a retrieval-augmented QA pipeline with fixed average chunk size 178 tokens, top-$$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$8, and generator $$o=[\mathrm{sim}(q,s_1),\dots,\mathrm{sim}(q,s_n)]\in[-1,1]^n$$9 Qwen2.5-7B, over CRUD, OmniEval, MultiFieldQA_zh, and HChemSafety. The reported QA metrics are BLEU, ROUGE-L, and METEOR (Zhao et al., 12 Mar 2026). Taken together, the two QChunker lines therefore cover both direct chunk scoring and end-to-end answer-quality evaluation.

5. Reported empirical behavior

QASC reports the strongest fully specified quantitative gains among the named QChunker variants. On the main comparison against fixed chunking at five granularities, recursive splitting, semantic chunking, and agentic chunking, the proposed method achieves Precision $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$0, Recall $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$1, F1-score $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$2, Answer Correctness $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$3, Faithfulness $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$4, and Latency $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$5 ms. The paper states that F1 $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$6 corresponds to a relative improvement of $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$7 over fixed chunking and $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$8 over semantic and agentic alternatives. Human ratings place QASC at 4.4 for Relevance, 4.2 for Coherence, 4.3 for Completeness, and 4.3 for Answer Quality (Rastogi, 29 Apr 2026).

System	Reported headline result	Additional evidence
QASC / “QChunker”	F1 $$t=\mathrm{Percentile}(o,p), \qquad S_{\mathrm{seed}}=\{s_i\mid \mathrm{sim}(q,s_i)\ge t\}.$$9	Faithfulness $p=75$0, Latency $p=75$1 ms
QChunker-3B	Outperforms all methods on all 12 metrics	HChemSafety: $p=75$2 BL, $p=75$3 ROU

QASC’s gains are also stable under several stratifications. By query type, it improves all four categories, “most dramatically on factoid (+9 ppt)” and by $p=75$4 ppt even for multi-hop queries. By document complexity, the relative gain over the best query-agnostic method remains a consistent $p=75$5 F1 points for low, medium, and high complexity documents. The ablation study attributes $p=75$6 ppt to the seed percentile strategy over top-$p=75$7 selection, $p=75$8 ppt to the contextual window ($p=75$9 versus $25\%$0), $25\%$1 ppt to weighted aggregation ($25\%$2 versus unweighted), and $25\%$3 ppt to merging and boundary adjustment (Rastogi, 29 Apr 2026).

The multi-agent QChunker reports its main results in comparative rather than fully enumerated table form. Across four heterogeneous domains and 12 task metrics, QChunker-3B outperforms rule-based, semantic, LLM-driven, and larger-model baselines. An ablation without $25\%$4 “drops consistently,” confirming the value of knowledge completion, and on HChemSafety the gains are described as especially pronounced, including $25\%$5 BL and $25\%$6 ROU. Additional analysis shows that knowledge completion lowers mean perplexity and removes “spikes” at context breaks for both 1.5B and 7B LLMs (Zhao et al., 12 Mar 2026).

The latency picture differs across the two formulations. QASC explicitly frames query-time chunking as a trade-off: it introduces approximately $25\%$7 ms of latency versus approximately $25\%$8 ms for fixed chunks, but the paper notes that precomputing sentence embeddings offline can save approximately $25\%$9 ms, and also proposes a hybrid use case in which QASC is applied only as a re-chunking step on the top-$s_r$0 retrieved documents (Rastogi, 29 Apr 2026).

6. Relation to broader chunking research, domain dependence, and limitations

The broader 2026 chunking literature provides the context in which QChunker should be interpreted. In retrieval-augmented code completion, a controlled empirical study across 864 settings shows that chunking strategy has a statistically significant effect; Function chunking underperforms the best strategy by 3.96 percentage points on RepoEval API-level EM and 5.64 percentage points on line-level EM, while Sliding Window and cAST dominate the cost–quality Pareto front and Function chunking is never Pareto-optimal. The same study identifies cross-file context length as the dominant parameter, with doubling from 2,048 to 8,192 tokens yielding up to $s_r$1 percentage points of improvement, whereas chunk size has a weaker, non-monotonic effect (Wu et al., 6 May 2026).

In question answering, a systematic industrial evaluation on Natural Questions finds that overlap provides no measurable benefit, sentence chunking is the most cost-effective method, and a “context cliff” reduces quality beyond approximately 2.5k tokens. In legal retrieval over the German Civil Code, section- and subsection-level chunking achieve the highest Recall@10 values, 0.46 and 0.47 respectively, while more complex approaches such as contextual chunking, RAPTOR, semantic clustering, and Lumber-style chunking perform worse and incur substantially higher offline build costs. A six-domain dense-retrieval benchmark similarly shows strong domain dependence: Paragraph Group Chunking is best overall with mean nDCG@5 of approximately 0.459, while Dynamic Token Size Chunking is strongest in biology, physics, and health (Bennani et al., 20 Jan 2026, Prior et al., 19 May 2026, Shaukat et al., 7 Mar 2026).

These results complicate a common assumption that more adaptive or more LLM-intensive chunking is always preferable. QASC shows large gains over fixed, semantic, and agentic alternatives on technical-document RAG, while legal retrieval favors chunk boundaries aligned with statutory structure, and code completion favors simple Sliding Window or cAST rather than function-level splits. A plausible implication is that QChunker methods are most compelling when the failure mode is specifically query misalignment, missing local context, or incomplete chunk self-sufficiency, rather than when the corpus already exposes strong native structure.

The limitations reported by the QChunker papers are correspondingly specific. QASC notes “distributed relevance,” vocabulary mismatch for rare technical terms, and implicit relevance as unresolved problems; its proposed future directions include relaxing contiguity, integrating neural query expansion such as HyDE, and adding cross-sentence coreference or graph-based extensions. The multi-agent QChunker notes that the pipeline still depends on a large LLM for initial annotation, uses only intra-document context, and may miss very long-range coherence breaks because of candidate sampling heuristics (Rastogi, 29 Apr 2026, Zhao et al., 12 Mar 2026). Together, these caveats place QChunker within a broader research program on chunking as an active optimization target rather than a fixed preprocessing routine.