PaperlayerView exact revision on arXiv

Source-linked AI summary

EnSI-RAG: Entity-Structure-Indexed Retrieval-Augmented Generation for Long-Document Question Answering

Xuanyu Meng, Jiashuo Sun, Jash Rajesh Parekh, Jiawei Han

arXiv:2608.21252v1cs.CLcs.AIcs.DBcs.IR

TL;DR

Long-document question answering is difficult when evidence is dispersed across passages and answers require multi-hop reasoning. EnSI-RAG indexes entity-centered evidence to guide retrieval, and reports competitive or superior performance against strong retrieval baselines across Loong and Oolong.

  • Problem

    Long-document QA remains challenging because evidence is dispersed across passages, while existing RAG places answer quality on retrieving evidence needed for aggregation and multi-hop reasoning.

    Source: Introduction · p. 1

  • Method

    EnSI-RAG builds a query-independent index of entity-centered passages and structured entity, type, property, relation, and aspect records that guide retrieval, while the LLM synthesizes answers from retrieved passages.

    Source: Introduction · pp. 1–2

  • Results

    Across Loong and Oolong, EnSI-RAG achieves competitive or superior performance compared with strong retrieval baselines.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    EnSI-RAG provides a flexible middle ground between unstructured chunk-based RAG and fully structured database-style reasoning across long-document QA settings.

    Source: Conclusion · p. 7

    Records need not encode every detail required by future questions; they serve as passage-addressing keys to original evidence rather than answer-ready reasoning state.

    Source: Information Extraction · p. 4

Abstract

from arXiv · show

Question answering (QA) over long, connected documents remains challenging because relevant evidence may span multiple entities and their relationships. Existing retrieval-augmented generation (RAG) methods typically index documents as raw chunks and retrieve them through embedding similarity. Their performance degrades when chunk boundaries separate entities from supporting evidence or when a question requires multi-hop reasoning across the corpus. We propose EnSI-RAG (Entity-Structure-Indexed Retrieval-Augmented Generation), a framework that constructs a query-independent, entity-centered index. Each record (e, t, k, v) represents an entity e, its type t, a semantic category k in {property, relation, aspect}, and a value v, while retaining links to the original source passages. At query time, these records serve as retrieval handles, and an LLM synthesizes the retrieved passages into the final answer. This design separates evidence localization from answer synthesis while preserving traceable source evidence. Across Loong and Oolong, EnSI-RAG achieves an average accuracy of 78.24. Relative to the published baseline scores used as references, this is 6.62 points higher, suggesting its effectiveness across these settings. The code is available at https://github.com/RamonMeng/EnSI-RAG.

Introduction

Long-document question answering requires locating evidence dispersed across entities and relationships, but conventional retrieval can fragment that evidence. EnSI-RAG uses entity-centered passages and a query-independent index to direct LLM synthesis to source passages.

Introduction

  • Long-document QA often requires evidence spread across passages, so systems must preserve entity identity, document structure, and relationships while locating relevant evidence.Applications include comparing financial reports, aggregating legal evidence, and tracing citations, methods, or findings across scientific papers.Source: Introduction · p. 1
  • Longer context windows do not ensure reliable evidence use, and RAG shifts answer-quality limits to the evidence retrieved.Models exhibit the “lost in the middle” phenomenon, while benchmarks find persistent difficulty with evidence aggregation and multi-hop reasoning (Liu et al. 2024; Bertsch et al. 2025). RAG reduces text processed by the LLM but makes retrieval quality a bottleneck (Lewis et al. 2020).Source: Introduction · p. 1
  • Fixed-length chunking can separate entities from evidence, detach table rows from headers, or combine unrelated topics, causing relevant evidence to be missed.Similarity-based retrieval may return locally relevant passages that are insufficient to answer a question or overlook semantically essential evidence with little lexical overlap.Source: Introduction · p. 1
  • EnSI-RAG defines retrieval units as semantically coherent, entity-centered passages rather than arbitrary chunks, reducing semantic fragmentation and contamination.These passages focus primarily on one entity and its local context, but require a query-independent representation to expose their semantic structure for future questions.Source: Introduction · p. 1
  • EnSI-RAG indexes entities, semantic types, properties, relations, and aspects as retrieval handles linked to original passages, while the LLM synthesizes retrieved evidence into answers.Structure improves evidence localization rather than replacing document-grounded reasoning with reasoning over normalized databases.Source: Introduction · pp. 1–2
  • EnSI-RAG is evaluated on Loong and Oolong, covering heterogeneous multi-document reasoning and large-scale information aggregation, respectively, with a unified framework and domain-specific passage construction.The introduction reports competitive or superior performance against strong retrieval baselines and describes the approach as a middle ground between chunk-based RAG and fully structured QA.Source: Introduction · p. 2

Related Work

Related work describes limitations in long-document evidence retrieval and in corpus representations that may discard source information. EnSI-RAG uses structured records to locate original passages for answer generation.

Related Work

  • EnSI-RAG uses structured records as retrieval handles to original evidence passages, avoiding the loss of source evidence that can accompany offline corpus structuring.Prior approaches include classical indexes, entity and relation extraction (Etzioni et al. 2004, 2008; Dong et al. 2014), and LLM-generated summaries, proposition spans, tables, schemas, and graphs (Sarthi et al. 2024; Chen et al. 2023; Chen et al. 2024; Jiang, Ma, and Chen 2024; Li et al. 2024; Liang et al. 2026; Wu, Zhang, and Li 2022); such structures can discard information or transfer poorly across question classes.Source: Related Work · p. 2
  • Larger context windows do not resolve long-document QA, as lost-in-the-middle effects hinder evidence use and aggregation-intensive questions (Du et al. 2025; Liu et al. 2024; Bertsch et al. 2025).Source: Related Work · p. 2
  • RAG grounds generation in retrieved evidence, and adaptive, iterative, and RL-trained retrieval policies improve this process, but still rely on near-perfect retriever recall (Lewis et al. 2020; Guu et al. 2020; Izacard et al. 2023; Asai et al. 2024; Trivedi et al. 2023; Jiang et al. 2023; Jin et al. 2025).Source: Related Work · p. 2
  • Fixed-size lexical or dense chunk retrieval can sever discussions, detach table rows from headers, merge unrelated entities, or omit evidence with little lexical overlap (Karpukhin et al. 2020; Izacard et al. 2022; Khattab and Zaharia 2020; Gong et al. 2020; Shin et al. 2025).Graph-based RAG models entity and relation structures, but is mainly developed and evaluated on short-passage benchmarks with well-canonicalized entities, assumptions that may fail for long, heterogeneous, domain-specific documents (Edge et al. 2024; Jiménez Gutiérrez et al. 2024; Parekh, Jiang, and Han 2025; Guo et al. 2024).Source: Related Work · p. 2

EnSI-RAG

EnSI-RAG builds a query-independent, entity-centered index over long-document passages, then uses it to locate source evidence for an LLM to synthesize into answers.

EnSI-RAG

  • EnSI-RAG separates offline passage indexing from online question answering, reusing the indexed corpus for future queries while generating answers from retrieved original passages.Its workflow constructs passages, extracts entity information, builds an index, retrieves evidence, and provides that evidence to an LLM.Source: EnSI-RAG · p. 2

Passage Construction

  • Entity-centered passages preserve an entity with supporting evidence while limiting unrelated content, with boundaries adapted to the task and document structure rather than fixed length.Each passage remains linked to its source through a unique psg_id, and the original text supports grounded retrieval and generation.Source: Passage Construction · pp. 2–3

Information Extraction

  • Extraction creates passage-supported retrieval records for entity properties, relations, and aspects, using records to locate evidence rather than to encode complete answers.Even a partial record can be useful if it leads to the correct source passage; unsupported information is excluded.Source: Information Extraction · p. 4

Index Entry Building

  • The index maps entity-structure keys to sets of supporting passages, preserving alternatives for lookup, comparison, aggregation, or multi-hop reasoning.Symbolic fields enable constrained matching, while textual representations support semantic matching across surface forms.Source: Index Entry Building · p. 4

Retrieval

  • Retrieval matches structured entity-centered keys to source passages and can use entities or values discovered in one hop to formulate later hops.Key ranking identifies compatible entity structures, while passage ranking selects useful source contexts among passages reached through those structures.Source: Retrieval · pp. 4–5

Generation

  • The LLM synthesizes answers from retrieved original passages, while the index records serve only as retrieval handles rather than a reasoning substrate.This separates evidence localization from answer synthesis and keeps the retrieved passages as the source of evidence.Source: Generation · p. 5

Experiments

Experiments evaluate EnSI-RAG on Loong and Oolong, compare its benchmark accuracy with published baselines, and examine design choices and per-query efficiency. Ablations favor row-level passages and domain-dependent label granularity, while efficiency results show substantial corpus-processing costs are offline.

Experimental Setup

  • Loong tests heterogeneous multi-document reasoning, while Oolong tests large-scale information aggregation over long inputs.The evaluation follows each benchmark’s official protocol: Loong uses average accuracy; Oolong uses an LLM judge for non-numeric questions and a deviation-based metric for numeric aggregation.Source: Experimental Setup · p. 5

Main Results

  • Table 2 compares EnSI-RAG accuracy with SLIDERS-reported baselines on Loong and Oolong; Loong results are references, not controlled head-to-head comparisons.Oolong uses the same deterministic benchmark slice, and each EnSI-RAG result comes from one complete run.Source: Main Results · p. 5, Table 2 · p. 6

Ablation Study

  • 100.00 versus 91.50: row-level passages outperform table-level passages on English Financial questions, so the study selects row-level passages for Financial documents.The authors attribute the gain to compact evidence units that reduce irrelevant context and give retrieval and generation a more precise interface.Source: Ablation Study · p. 6, Table 4 · p. 7
  • Coarse labels raise English Financial accuracy from 92.00 to 93.50 and Paper from 88.00 to 91.50, but lower Legal accuracy from 71.25 to 68.75.Property and aspect names are mapped to higher-level concepts, while directional relation labels are preserved; the study selects coarse labels for English Financial and Paper and fine labels for Legal.Source: Ablation Study · pp. 6–7, Figure 3 · p. 7
  • 94.00: Top-5 has the highest English Financial accuracy, above Top-15 at 92.50 and Top-12 at 92.00.Despite this subset result, the main experiments use Top-12 for a common configuration and to avoid benchmark-specific tuning.Source: Ablation Study · p. 7, Figure 4 · p. 7

Efficiency

  • 163,146 input tokens, 494.30 seconds, and 34.35 LLM calls per query: EnSI-RAG performs costly corpus processing offline and limits online work to retrieval and generation.After amortization, Passage Construction and Information Extraction account for 157,031 input tokens, 468.51 seconds, and 32.35 calls; online Retrieval uses 656 tokens, 19.01 seconds, and 1 call, while Generation uses 5,460 tokens, 6.73 seconds, and 1 call. Index Entry Building adds 0 input tokens and calls and 0.05 seconds.Source: Table 3 · p. 6, Efficiency · p. 7

Conclusion

EnSI-RAG replaces arbitrary fixed-size chunk indexing with a query-independent, entity-structure index linked to original supporting passages. It separates evidence localization from answer synthesis, with an LLM integrating the retrieved passages.

Conclusion

  • EnSI-RAG builds a query-independent index connecting entities, their types, and properties, relations, or aspects to semantically contained passages.The index maps each structured record to original passage IDs.Source: Conclusion · p. 7
  • The index provides structured paths from future questions to original supporting passages, improving evidence localization while leaving final semantic integration to the LLM.Source: Conclusion · p. 7
  • EnSI-RAG is designed as a flexible middle ground between unstructured chunk-based RAG and database-style reasoning across financial, legal, academic, and aggregation-heavy QA settings.Source: Conclusion · p. 7

Appendix: Experimental Details

The appendix describes the benchmarks, their evaluation protocols, and the baseline methods used in the experiments.

Appendix: Experimental Details

  • Table A1 describes the benchmarks; evaluated subset sizes and task types are reported in the main paper.The appendix passage does not provide those subset sizes or task types.Source: Table A1 · p. 10
  • Evaluation follows each benchmark’s official protocol, and selected baselines span six methodological families.These families are conventional retrieval-augmented generation, long-context inference, document transformation, agent-based reasoning, recursive reasoning, and structured retrieval.Source: Table A2 · p. 11, Table A3 · p. 12

Case Studies

Two case studies show how preprocessing structure affects whether EnSI-RAG can recover complete answers: legal matching depends on preserving benchmark-facing identifiers, while citation-chain construction depends on complete citation links and canonical titles.

Legal judgment-result matching

  • In legal judgment-result matching, the task requires a one-to-one JSON mapping from each benchmark document ID to its option ID, not a summary or title-level answer.The question supplies 12 legal documents and 15 candidate results; keys must be benchmark-local document labels and values option identifiers.Source: Table A4 · p. 13, Case Studies · p. 13
  • SLIDERS captures legal attributes but omits the benchmark document label as the task-facing key, producing summaries, frequent-result predictions, or title-level mappings instead of the required mapping.The outputs may describe cases correctly but do not provide the required Doc-i to Result-j mapping.Source: Case Studies · p. 13

Paper citation-chain construction

  • Citation-chain construction requires the longest valid directed path across five supplied papers, returned in chronological citation order with canonical titles and without non-linear branches.The gold chain runs from “Towards Understanding Chain-of-Thought Prompting: An Empirical Study of What Matters” to “ChatGPT believes it is conscious.”Source: Table A5 · p. 14, Case Studies · p. 14
  • SLIDERS extracts only three of the four citation edges needed for the gold chain and leaves some titles noncanonical, so its response omits a required paper and is incorrect.The missing edge is from “Capabilities of GPT-4 on Medical Challenge Problems” to “Large Language Models Encode Clinical Knowledge”; once absent, the complete path cannot be reconstructed from the reconciled table.Source: Case Studies · p. 14

Additional Reproducibility Details

The Oolong evaluation uses a verified deterministic 192-question test slice, and the paper reports model settings and one complete run per configuration. It does not report across-run variability or statistical significance tests.

Oolong slice verification

  • The Oolong slice contains 192 eligible test questions selected deterministically after filtering for 262,144-token documents.The program verifies the split, count, document lengths, and reconstructed question strings; mismatches stop evaluation without a score. Selected question identifiers are released.Source: Additional Reproducibility Details · p. 14, Table A6 · p. 15

API and local execution configuration

  • GPT-4.1-mini preprocesses with LLMs, while GPT-4.1 handles retrieval planning and answer generation; all API calls use temperature 0.Loong uses Top-12 retrieval and octen-embedding-8B for its Financial dataset; Oolong-Synth selects records deterministically without an embedding model. API-side hardware and serving configuration are unavailable.Source: Additional Reproducibility Details · p. 15

Number of runs

Prompts

The appendix reproduces the complete prompts used in the Loong and Oolong experiments; deterministic pipeline stages that do not invoke an LLM have no prompts to report.

Prompts

  • The appendix provides the complete prompts used in the Loong and Oolong experiments, reproduced verbatim from the experiment scripts.Braced expressions denote runtime substitutions in the corresponding templates.Source: Prompts · p. 16
  • Deterministic parsing, index construction, filtering, and score aggregation do not invoke an LLM and therefore have no prompts to report.Source: Prompts · p. 16

Related papers

Loading 2608.21252v1…