NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor

Long Wang

cs.IROct 8, 2026

NativeScope: Relation-Localized Retrieval over Native Topology with a Correct Anchor

Authors: Long Wang

Abstract

Dense retrieval usually ranks text chunks by their semantic similarity to a question. This ignores structure that many data systems already store, including section membership, session boundaries, and native order. We propose NativeScope, a scope-then-rank method for queries with a known anchor and relation. It represents a query as q -> (A, r, B). The anchor A and relation r select native units through belonging, before, or after operators, and the target term B ranks only chunks that overlap the selected scope. An internal variant, NS-FullQ, ranks the same candidates with the full question. We evaluate both methods on 200 controlled document and memory records derived from QASPER and LongMemEval under a 1,024-token budget. NativeScope attains native-unit recall of 89.28 percent for documents and 72.50 percent for memories, improving over instance-wide Dense RAG by 42.75 and 22.00 percentage points. NS-FullQ reaches 87.78 percent and 68.50 percent; its differences from NativeScope are inconclusive, locating the primary gain in relational scoping rather than the shorter ranking query. With automatic Top-1 anchors, memory recall falls to 35.50 percent. NativeScope is therefore effective when anchor coordinates and native relations are reliable, but hard scoping inherits errors from the localization interface.

Figures & tables

Figure 1: NativeScope pipeline. A decomposed query and correct anchor drive a native relation operator. Eligible units are mapped to candidate text chunks by character-interval overlap. NativeScope and NS-FullQ share the same scoping stage and differ only in within-scope ranking input. Gold target locations and answers are not exposed to the ranker.

Figure 2: Native relation operators and unit-to-chunk mapping. Belonging selects the anchor’s container, while before and after select up to eight units in the respective direction under global native order. Colored bars denote chunks that may span several native units.

Method	Candidate scope	Ranking input/basis	Role
Dense RAG	Full instance	Full question $q$	Flat dense baseline
Anchor Window	Anchor $\pm 8$ units	Native distance	Distance-window control
NS-FullQ	Chunks overlapping $\Gamma_{r}(a,T_{X})$	Full question $q$	Internal ranking variant
NativeScope	Same as above	Target term $B$	Target-ranking implementation

Table 1: Interfaces of the main comparisons. The last three methods receive the same correct anchor; all methods share chunks, encoder, and return budget.

Method	Document	Memory
Dense RAG	46.53	50.50
Anchor Window	45.36	2.00
NS-FullQ (internal variant)	87.78	68.50
NativeScope	89.28	72.50

Table 2:

\mathrm{Recall}@1024

with correct anchors (percent). Each domain has

N=100

under fixed 256-token chunking. Bold marks only the largest point estimate.

Figure 3: Two-domain main results under correct anchors (

N=100

per domain). The vertical axis is gold native-unit Recall@1024; all methods use the same chunks, encoder, and return budget.

Domain	Control	Difference	95% interval
Document	Dense RAG	$\boldsymbol{+42.75}$	$\boldsymbol{[33.25,52.50]}$
Document	Anchor Window	$\boldsymbol{+43.92}$	$\boldsymbol{[34.75,53.34]}$
Document	NS-FullQ	$+1.50$	$[-2.00,5.50]$
Memory	Dense RAG	$\boldsymbol{+22.00}$	$\boldsymbol{[12.00,32.00]}$
Memory	Anchor Window	$\boldsymbol{+70.50}$	$\boldsymbol{[61.50,79.00]}$
Memory	NS-FullQ	$+4.00$	$[-3.00,11.00]$

Table 3: Paired NativeScope-minus-control differences and 95% intervals, in percentage points. Bold marks comparisons whose interval lies entirely above zero; no advantage is claimed when an interval includes zero.

Domain	Relation	$N$	Dense	Window	NS-FullQ	NativeScope
Document	belonging	50	46.06	53.56	86.06	92.06
Document	before	25	55.00	37.00	92.00	86.00
Document	after	25	39.00	37.33	87.00	87.00
Memory	belonging	50	39.00	2.00	53.00	61.00
Memory	before	25	60.00	4.00	80.00	84.00
Memory	after	25	64.00	0.00	88.00	84.00

Table 4:

\mathrm{Recall}@1024

by relation under correct anchors (percent). These groups partition the main evaluation. Bold marks the largest point estimate in each row; ties are included.

Metric/method	Document	Memory
Anchor Top-1 accuracy	61.00	37.00
All gold targets in relation scope	74.00	53.00
Dense RAG recall	46.53	50.50
Anchor Window recall	43.69	3.00
NS-FullQ recall	65.44	33.50
NativeScope recall	66.94	35.50

Table 5: Default-chunking results under the original automatic-localization protocol (percent). Localization and scope coverage are diagnostics for the relation-localized pipeline. Within the retrieval-method rows, bold marks the largest recall in each domain.

Figure 4: Effect of the anchor condition on relation-localized recall. Correct-anchor and original automatic-localization results share default chunking and Recall@1024, while the main conditional evaluation was introduced after the automatic degradation was observed.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Batch	Results	Anchor condition	Use in this paper
Original chunking main run	4,000	Automatic	Default-chunking boundary result; other runs retained as exploratory records
Original mechanism batch	9,280	Automatic and separate diagnostics	Audit retained; not evidence for mechanisms under the revised condition
Correct-anchor revision	800	Given correct coordinate	Main evidence in Tables 2 – 4
Generative QA	36	Original protocol, incomplete	Paused; no method comparison reported

Appendix

Table 6: Executed material and its evidential role. Completed runs do not establish all earlier claims.

Explore similar work

ShardMemo: Scope-Before-Routing for Agentic Memory Retrieval
Jan 29, 2026Yang Zhao, Chengxiao Dai, Mengying Kou +2Agentic Retrieval Information Retrieval
Anonymous
Query Expansion Should Be Coordinated: Dense Expands, Sparse Anchors
Aug 16, 2026Chunran ZhangRetrieval-Augmented Generation RAG Evaluation
School of Computing and Artificial Intelligence Southwest Jiaotong University Chengdu, China
MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval
Jun 16, 2026Amirhossein Abaskohi, Raymond Li, Gaetano Cimino +3Retrieval-Augmented Generation Information Retrieval
University of British Columbia · University of Salerno · ServiceNow Research

Jan 29, 2026, cs.AIJ/K move · Enter open · S save

ShardMemo: Scope-Before-Routing for Agentic Memory Retrieval

Yang Zhao, Chengxiao Dai, Mengying Kou, Yue Xiu+1

Anonymous

Agentic Retrieval Information Retrieval Agent Memory Management

Open paper AskPDF arXiv

Agentic systems accumulate persistent memory across sessions, tools, and tasks, and a later request must retrieve from it under two distinct constraints: which memories it is permitted to access, and which are relevant under a limited search budget. Existing memory systems handle access scope in two flawed ways: applying scope after retrieval wastes probe budget on inadmissible memories, while treating scope as a learned ranking feature makes a hard constraint depend on router quality. We present SHARD-MEMO, an agentic memory system built on scope-before-routing: metadata predicates first identify the admissible shards, and a learned router then selects a small number of them for shard-local approximate nearest neighbor retrieval. Given the supplied scope predicate and metadata, this separates hard admissibility from learned relevance ranking, so inadmissible shards cannot consume shard-probe budget. We evaluate on LoCoMo, HotpotQA, and ToolBench, covering conversational, long-context, and procedural memory. Under matched supervision and fixed budgets, SHARDMEMO improves over a learned router baseline by roughly +3 F1 on LoCoMo; in end-to-end LoCoMo evaluation it improves over the strongest evaluated memory baseline by up to +6.8 F1, with gains on HotpotQA and ToolBench.

Abstract

Explore similar work

ShardMemo: Scope-Before-Routing for Agentic Memory Retrieval

Query Expansion Should Be Coordinated: Dense Expands, Sparse Anchors

MCompassRAG: Topic Metadata as a Semantic Compass for Paragraph-Level Retrieval