EAR: Entity-Aware Partitioning Approach for Retrieval-Augmented Generation Development

• EAR extracts normalized surface anchors from questions, options, and corpus text, then retrieves compact windows around matching corpus anchors.The method makes the retrieval unit responsive to the MCQA item rather than only to document segmentation.Source: §I · p. 1
• The study asks how accuracy and context cost change when generic chunks are replaced with entity-aware windows.Source: §I · p. 1
• The current surface-anchor rules are tailored to one textbook curriculum's language and subject, so cross-domain transfer is not established.Source: §I · p. 1
• The authors introduce an auditable rule-based schema, entity-window retrieval, and an optional parent-expanded variant.Source: §I · p. 1
• The evaluation compares no-RAG, chunk, entity-window, and parent-expanded approaches across two local and one hosted model on 153 supported questions.Direct comparisons hold the split, model, prompt family, and cleaned corpus fixed.Source: §I · p. 1
• Reported measures include accuracy, paired uncertainty, context volume, anchor coverage, computational cost, and window-radius sensitivity.Source: §I · p. 1

• Fixed chunks are common RAG units, while longer contexts can make relevant evidence harder for language models to use reliably.Related approaches include Self-RAG passage assessment and specialized clinical entity retrieval.Source: §II · p. 1
• GraphRAG uses entity graphs and community summaries, [11], whereas EAR uses normalized surface anchors to select local windows; parent context is tested as an optional expansion.Source: §II · p. 2
• Prior Turkish RAG studies compare retrieval and generation models, pipeline stages, or textbook systems,,,; EAR asks whether deterministic surface-anchor windows suit MCQA.RAGTurk reports added generative modules can increase cost without stable gains.Source: §II · p. 2

A. Task and Evaluation Protocol

• The task is five-way MCQA, and direct entity-window versus chunk comparisons hold the generator, prompt, split, corpus, and top-k setting fixed.Parent-expanded retrieval tests broader context; BM25 pages appear only in the Mistral diagnostic.Source: A. Task and Evaluation Protocol · p. 2

B. Data Cleaning and Corpus Construction

• An automatic support heuristic reduced the strict 672-question test split to 153 questions, but its support judgments lack completed human validation.Source: B. Data Cleaning and Corpus Construction · p. 2
• The corpus uses cleaned grade 9–12 curriculum textbooks, with exercises and multiple-choice blocks removed; Table I summarizes the evaluated data regime.The audit reports 1,108 pages, 3,955 kept blocks, and 4,429 dropped blocks.Source: B. Data Cleaning and Corpus Construction · p. 2, Table I · p. 2

• EAR defines an entity as a normalized textual anchor, extracted from years, capitalized or event-like phrases, and short answer options.Normalization handles locale-aware case folding, punctuation, numerals, whitespace, and domain suffixes.Source: C. Proposed Method: EAR · p. 2
• The method scores windows by lexical overlap plus anchor and year/date bonuses, falls back to lexical scoring when no anchor matches, and removes duplicate windows.EAR requires neither an ontology nor a learned linker or embedding model.Source: C. Proposed Method: EAR · p. 2
• The corpus pipeline segments cleaned pages into overlapping parent chunks, extracts anchors, and indexes child windows around them; inference retrieves the top-k deduplicated windows.Child windows extend up to 80 words on each side of an anchor.Source: C. Proposed Method: EAR · p. 2
• Algorithm 2 expands Algorithm 1's entity-window output with parent context and an extractive summary.Figure 1 identifies Algorithm 1 as entity-window retrieval and Algorithm 2 as the parent-expanded variant.Source: C. Proposed Method: EAR · p. 2, Figure 1 · p. 3
• 98.7% of supported items had an exact normalized corpus-index anchor match, while 50.8% of extracted anchors matched an index key.These are operational coverage measures, not NER precision or recall; no manually labeled entity set was available.Source: C. Proposed Method: EAR · p. 2

D. Design Rationale and Scope

• EAR ranks local windows around searchable anchors rather than ranking fixed passages against the whole query.The pipeline identifies anchors, then ranks windows around corpus matches.Source: D. Design Rationale and Scope · p. 3
• Each returned context can be traced to a query anchor, corpus anchor, and window, separating extraction, retrieval, and generation errors for inspection.The stages are query extraction at Step 5, retrieval at Steps 6–7, and generation after evidence is returned.Source: D. Design Rationale and Scope · p. 3
• The current capitalization, suffix, and normalization rules are corpus-specific, so transfer requires redesigning and validating query and corpus extraction.The study does not establish cross-domain generalization.Source: D. Design Rationale and Scope · p. 3

E. Baselines, Prompts, and Answering Protocol

• The chunk baseline uses overlapping 220-word passages with 30-word overlap, while BM25 page retrieval is included only as a diagnostic.Chunks are ranked by lexical token overlap; BM25 page retrieval applies query-aware evidence compression.Source: E. Baselines, Prompts, and Answering… · p. 3
• Retrieval runs use RAG+CoT, while direct, few-shot, and CoT prompts without retrieval provide the no-RAG baselines.The sweep evaluates three named models at top-k = 3 and 8 with temperature 0 and one generation per item.Source: E. Baselines, Prompts, and Answering… · p. 3

F. Metrics and Paired Tests

• End-to-end accuracy on the supported split is the primary metric, with invalid-answer rate, retrieved words, latency, and paired accuracy differences as secondary measures.Outputs are parsed as A–E, with malformed outputs marked invalid.Source: F. Metrics and Paired Tests · p. 3
• Paired correctness is tested with exact McNemar tests, and 95% confidence intervals use a 2000-resample percentile bootstrap.The bootstrap uses seed 13.Source: F. Metrics and Paired Tests · p. 3
• Absolute scores measure end-to-end retrieval QA, not closed-book model knowledge, because the task requires retrieval, option selection, supported evidence, and strict parsing.The relevant comparison is accuracy and retrieved-context length under the same protocol.Source: F. Metrics and Paired Tests · p. 3
• All model runs cover 153 items, while separate diagnostics measure index and retrieval costs and compare 40-, 80-, and 120-word radii.Gold-option surface-form presence is logged independently from correctness and is a cue diagnostic, not proof of answer support.Source: F. Metrics and Paired Tests · p. 3

IV. RESULTS

• 37.5% fewer words at top-k = 3 and 40.2% fewer at top-k = 8: EAR entity-window retrieves fewer words than chunks, while accuracy changes vary by model.At top-k = 3, accuracy changes are +5.2, +1.3, and -3.9 points for Mistral, Gemma, and DeepSeek; at top-k = 8, they are +5.9, -3.3, and -4.6 points.Source: A. Finding 1: Context Savings Are… · p. 3
• The observed retrieval results do not improve accuracy in the high direct-baseline regime.DeepSeek reaches 83.0% directly, and no retrieval row exceeds it.Source: A. Finding 1: Context Savings Are… · p. 3
• Every entity-window confidence interval crosses zero, and no entity or parent comparison remains significant after Holm or Bonferroni correction.The comparisons cover 12 entity and parent results.Source: A. Finding 1: Context Savings Are… · pp. 3–4

B. Finding 2: Compact Windows Shift Offline and Online Cost

• 24.4–25.2 ms per item: anchor-filtered entity retrieval is faster than exhaustive chunk scoring at 60.6–62.4 ms, despite higher offline indexing cost.EAR builds 33,228 windows in 6.877 s and 143.0 MiB of peak traced Python allocation, versus 0.011 s and 3.4 MiB for 1397 chunks; entity-window end-to-end latency is lower by 1.8–31.4% across six runs.Source: B. Finding 2: Compact Windows Shift… · p. 4
• At top-k = 3, the 80-word radius raises gold-option surface-cue rate from .529 at 40 words to .608, while 120 words reaches .601 and retrieves 35.5% more text than 80.At top-k = 8, 120 words reaches .706 versus .673 for 80, while retrieving 1409.8 rather than 1062.3 words; 80 words is a fixed engineering choice, not accuracy-tuned.Source: B. Finding 2: Compact Windows Shift… · p. 4

C. Finding 3: Parent Expansion and Surface Cues Have Limits

• Parent expansion has the highest observed Mistral and Gemma accuracy at top-k = 3, but falls below entity-window for Mistral and chunks for Gemma and DeepSeek at top-k = 8.The suggested distractor mechanism from added narrative was not directly tested; the Gemma top-k = 8 parent-versus-chunk difference does not survive multiple-comparison correction.Source: C. Finding 3: Parent Expansion and… · p. 4
• BM25 page has the highest gold-option surface-cue rate (.882) in the Mistral top-k = 8 diagnostic, yet its accuracy (.431) is below entity-window (.444).A present cue is not sufficient evidence of support: entity-window has 44 wrong answers despite a cue, and human evidence annotation is needed to assess cue support.Source: C. Finding 3: Parent Expansion and… · p. 4

V. DISCUSSION

• EAR reduces retrieved text by 37.5–40.2% while moving work into a larger offline index; anchor filtering is faster than exhaustive chunk scoring, but parent expansion reverses that online advantage.The pattern supports evaluating pipeline complexity against cost rather than assuming benefit.Source: §V · p. 4
• Mistral’s entity-window point estimates are higher at both budgets, Gemma’s direction changes with k, and DeepSeek’s strong direct baseline is not improved by retrieval; none of the differences is statistically significant.The local-model scores reflect a restricted-corpus pipeline with shuffled-option parsing and dependence on generator use, not closed-book knowledge estimates.Source: §V · p. 4, Table II · p. 5
• BM25 pages contain the gold-option surface form more often than EAR windows in the Mistral diagnostic without producing higher accuracy, and logs expose possible failure stages from anchor extraction through generator use.Human passage judgments are needed to distinguish genuine supporting evidence from incidental answer-string presence.Source: §V · p. 4
• EAR is a partitioning component, not a complete QA architecture; broader evaluation should compare learned extractors, validate the supported split, and rerun the radius sweep end to end.Relation-aware expansion is proposed for questions about cause, chronology, institutional role, or comparison.Source: §V · p. 5

VI. LIMITATIONS

• The 153-question, three-generator evaluation does not establish broad model or domain trends; the supported split lacks completed human validation, and cost measurements are hardware-specific.The rule-based extractor has no manually labeled anchor set for precision, recall, or F1, and its capitalization and suffix rules can miss aliases, morphology, and cross-sentence references.Source: §VI · p. 5

VII. CONCLUSION

• 37.5–40.2% fewer words than chunks and mixed accuracy point estimates support EAR as a compact, inspectable retrieval representation, not a reliable default for parent expansion.No entity-window difference is statistically significant; extractor validity, corpus support, and cross-domain transfer remain open.Source: §VII · p. 5