Skip to content

Retrieval-Augmented Generation with Retrieval-optimized TRA Structure

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Tiabet/TRA-RAG
Area: Information Retrieval & RAG
Keywords: retrieval-augmented generation, knowledge graph, multi-hop reasoning, query decomposition, hybrid retrieval

TL;DR

Addressing the contextual loss and structural rigidity of triplet graphs in Graph-based RAG, this paper proposes an entity-centric hierarchical Title-Relations-Attributes (TRA) database structure and atomic Fact-level indexing, coupled with dependency-aware query decomposition and reciprocal rank fusion, achieving comprehensive state-of-the-art performance across simple and multi-hop QA benchmarks.

Background & Motivation

Retrieval-Augmented Generation (RAG) grounds Large Language Model (LLM) responses in external knowledge bases, substantially reducing hallucinations and knowledge obsolescence. To capture deeper relational paths beyond surface keyword matching or dense semantic similarity, graph-based RAG has attracted significant research interest. Conventional graph RAG systems prompt LLMs to extract (subject, predicate, object) triplets directly from an unstructured corpus and build knowledge graphs, relying on graph connectivity and multi-hop entity traversal to answer complex questions. However, converting natural language into atomic triplets incurs a severe information bottleneck: real-world text exhibits rich hierarchical relationships and n-ary dependencies that cannot be flattened into isolated edge pairs without contextual damage. For instance, structured facts such as "an actor played a specific role in a movie, which marked his first villain appearance" become fragmented across disconnected triplets, stripping the essential association between the role and the specific film context.

Existing alternatives fail to strike an optimal balance between granularity and contextual coherence. Using raw document passages or arbitrary text chunks as search units preserves raw narrative flow but introduces excessive noise and coarse retrieval boundaries. Conversely, tree-structured summarization methods (such as RAPTOR) obscure fine-grained numerical attributes and exact entity references during recursive compression. Furthermore, while extensive research focuses on sophisticated graph traversal algorithms and agentic routing, relatively few studies reconsider the foundational search target itself. When the underlying database representation is structurally rigid or fragmented, even the most elaborate graph search strategies suffer from semantic drift and high inference latency across long reasoning trajectories.

This work addresses this tension by re-anchoring corpus modeling in classical Entity-Relationship (ER) design principles and hierarchical metadata architectures. Rather than scattering facts across unmoored binary links, the framework structures all attributes and outbound relations around a central entity. Core idea: reconstruct raw text passages into hierarchical TRA (Title-Relations-Attributes) JSON structures and index atomic Fact units via pseudo-natural language serialization, enabling accurate and low-latency multi-hop reasoning through dependency-aware query decomposition and dual-track reciprocal rank fusion without expensive graph traversal.

Method

Overall Architecture

The TRA-RAG framework consists of an offline structural indexing pipeline and an online multi-stage retrieval and reasoning pipeline. In the offline phase, an LLM parses unstructured text passages into hierarchical JSON metadata following the TRA schema, defining the primary entity as Title, internal properties and nested events as Attributes, and external links as Relations. The system then segments the TRA tree into atomic Fact units defined by "Title + Top-level Key + full Value" and serializes them into pseudo-natural language sentences, which are indexed simultaneously via dense vector embeddings and sparse BM25 indices. During the online phase for complex multi-hop queries, a query decomposition module breaks the main query into a directed acyclic chain of sub-questions with explicit answer dependencies. A sequential solver retrieves relevant Facts for each sub-question via Reciprocal Rank Fusion (RRF) and generates intermediate answers conditioned on previous question-answer history. Finally, a dual-track reranker aggregates global core evidence and sub-query-specific clues to supply the final answer synthesis model. For simple single-hop queries, query decomposition and reranking are bypassed, executing direct hybrid retrieval.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Unstructured Text Corpus"] --> B["TRA Structure Construction<br/>Extract Hierarchical Title / Attributes / Relations JSON"]
    B --> C["Fact Atomic Indexing<br/>Flatten to Pseudo-Natural Sentences + Hybrid Dense/Sparse Indexing"]
    C --> D["Query Decomposition & Dependency Modeling<br/>Split Multi-Hop Query into Dependent Sub-Questions"]
    D --> E["Sequential Sub-Question Answering & Context Chaining<br/>Stepwise Fact Retrieval with Prior QA History"]
    E --> F["Dual-Track Evidence Reranking & Consolidation<br/>Merge Global Seed Evidence and Sub-Query Specific Clues"]
    F --> G["Final Answer Synthesis"]

Key Designs

1. TRA Structure Construction: Overcoming Triplet Rigidity via Entity-Centric Hierarchical Representation To eliminate the contextual fragmentation inherent in conventional graph-based RAG, the authors design the TRA (Title-Relations-Attributes) structure inspired by relational ER modeling and metadata organization. Encapsulated in a hierarchical JSON schema, each unit designates the central concept or person in a passage as Title, structures fine-grained descriptions, nested timelines, cast lists, and categorical details under Attributes, and groups outbound entity connections under Relations. Unlike rigid triplet extraction where multi-faceted associations are stripped into binary edges, TRA maintains relational cohesion within nested dictionaries (e.g., keeping an actor, character role, and award status inside the same movie attribute block), resolving pronouns naturally and avoiding cross-triplet information loss.

2. Fact Atomic Indexing: Retrieval-Optimized Search Space Balancing Detail and Coherence Embedding an entire complex JSON object directly degrades dense retrieval precision because off-the-shelf embedding models struggle with deep syntax formatting and broad multi-topic scopes, whereas fragmenting JSON down to individual key-value pairs recreates triplet-style semantic collapse. To establish an optimal search target, the authors define a Fact as an atomic retrieval unit comprising "Title + Top-level Key + entire Value". To accommodate embedding models, each Fact is flattened into pseudo-natural language through template serialization (e.g., "Key of Title is Sub-Key-1 is Sub-Value-1, Sub-Key-2 is Sub-Value-2..."), while relation items use "Title + Key + Value". Every serialized Fact is indexed into Fact Storage using dense vector embeddings (OpenAI text-embedding-3-small) and BM25 sparse lexical matching (with stopword removal), providing complementary semantic generalization and exact entity recall.

3. Dependency-Aware Query Decomposition & Sequential Answering: Mitigating Multi-Hop Semantic Drift Multi-hop queries inherently suffer from semantic divergence where intermediate reasoning steps bear little lexical or vector similarity to the original overarching query. Rather than navigating graphs via speculative entity walks, TRA-RAG prompts an LLM to decompose the complex question into an ordered sequence of single-hop sub-questions \(\{q_1, q_2, \dots, q_n\}\) with explicit dependency placeholders (e.g., [SQ1_Answer]). During the answering phase, the model solves each sub-question sequentially, retrieving relevant Facts and conditioning the \(i\)-th sub-answer \(a_i\) on the cumulative preceding dialogue history \(H_{i-1} = \{(q_1, a_1), \dots, (q_{i-1}, a_{i-1})\}\) and current retrieval context \(R_i\):

\[a_i = \text{LLM}(q_i \mid R_i, H_{i-1})\]

This chaining mechanism prevents semantic drift by providing concrete intermediate entity bindings at each step.

4. RRF Hybrid Retrieval and Dual-Track Reranking: Eliminating Score Domination and Evidence Loss In each retrieval stage, the system combines dense embedding ranks \(r_d(f)\) and BM25 sparse ranks \(r_{bm}(f)\) using Reciprocal Rank Fusion (RRF) with smoothing parameter \(k=60\), dense weight \(w_d=1.3\), and sparse weight \(w_{bm}=1.0\):

\[\text{Score}(f) = w_d \cdot \frac{1}{k + r_d(f)} + w_{bm} \cdot \frac{1}{k + r_{bm}(f)}\]

To prevent "score domination"โ€”where disproportionately high absolute retrieval scores from one sub-query crowd out subtle yet critical evidence from anotherโ€”the authors devise a dual-track context selection strategy:

\[R_{\text{final}} = \{F_{\text{seed}} \cup P_{\text{seed}}\} \cup \{F_{\text{re}} \cup P_{\text{re}}\}\]

The Global Relevance track aggregates all retrieved Facts across sub-questions by their RRF scores, extracting the top \(F_i\) Facts (\(F_{\text{seed}}\), default 20) and top \(P_i\) unique source passages (\(P_{\text{seed}}\), default 3). The Sub-Query Specificity track then performs secondary reranking on the remaining candidate pool using Min-Max normalized score fusion \(\text{Score}_{\text{re}}(f, q_{\text{sub}}) = w_d \cdot \hat{s}_d(f, q_{\text{sub}}) + w_{bm} \cdot \hat{s}_{bm}(f, q_{\text{sub}})\), selecting top localized Facts \(F_{\text{re}}\) (default 10) and passages \(P_{\text{re}}\) (default 2). This ensures that subtle entity clues unique to a single reasoning step are faithfully preserved alongside broad topic evidence.

A Worked Example

Consider the multi-hop query: "How long are the council terms of the city that shares a county with Helvetia?" 1. Query Decomposition: The query is decomposed into \(q_1\): "In which county is Helvetia located?", \(q_2\): "What city shares a county with [SQ1_Answer]?", and \(q_3\): "How long are the council terms in [SQ2_Answer]?". 2. First Hop Retrieval & Answering: Using \(q_1\), hybrid retrieval matches the Fact from the Helvetia entry: "County of Helvetia, Arizona is Pima County", returning \(a_1 = \text{Pima County}\). 3. Second Hop Dependency Resolution: Substituting \(a_1\) into \(q_2\) yields "What city shares a county with Pima County?". Hybrid retrieval identifies Tucson's Fact: "County seat of Tucson, Arizona is Pima County", yielding \(a_2 = \text{Tucson}\). 4. Third Hop Retrieval & Consolidation: Substituting \(a_2\) into \(q_3\) retrieves Tucson's administrative structure: "The government structure of Tucson, Arizona is council members term is four-year terms". Dual-track reranking aggregates the 5 source passages and 30 top Facts, allowing the final LLM synthesis step to verify the reasoning chain and correctly output: "four years".

Key Experimental Results

Main Results

Evaluation was conducted on two Simple QA datasets (NQ, PopQA) and four Multi-Hop QA datasets (HotpotQA, 2WikiMultiHopQA, MuSiQue, LV-Eval). All baselines and TRA-RAG standardize on gpt-4o-mini for generation and text-embedding-3-small for embeddings, using a fixed 5-passage budget for answer generation.

Group Method NQ (F1 / Acc) PopQA (F1 / Acc) HotpotQA (F1 / Acc) 2Wiki (F1 / Acc) MuSiQue (F1 / Acc) LV-Eval (F1 / Acc) Avg (F1 / Acc)
Non-Retrieval Pure-LLM 51.2 / 57.9 28.2 / 35.0 36.8 / 45.9 29.9 / 35.0 17.1 / 17.8 8.0 / 7.3 28.5 / 33.1
Non-Retrieval Oracle (100% Recall) 68.5 / 70.4 57.9 / 64.4 75.3 / 85.7 60.5 / 66.1 52.0 / 56.3 - / - 62.8 / 68.6
Naive Retrieval Cosine-Sim 58.7 / 60.1 59.7 / 68.1 56.2 / 65.1 39.3 / 42.7 24.5 / 26.2 2.9 / 0.8 40.2 / 43.8
Naive Retrieval Hybrid (RRF) 59.5 / 61.7 60.3 / 69.3 56.4 / 66.0 37.3 / 40.4 20.8 / 21.7 2.9 / 0.8 39.5 / 43.3
Structure-Augmented GraphRAG 39.0 / 42.0 39.5 / 48.8 60.8 / 72.0 43.5 / 47.4 24.7 / 27.3 14.0 / 12.1 36.9 / 41.6
Structure-Augmented LightRAG 37.0 / 40.0 33.5 / 42.8 42.1 / 50.8 40.4 / 44.3 21.7 / 26.5 12.4 / 11.3 31.2 / 35.9
Structure-Augmented RAPTOR 47.0 / 50.0 52.0 / 61.3 64.2 / 74.1 46.6 / 52.3 27.8 / 30.9 9.3 / 7.3 41.1 / 46.0
Structure-Augmented LinearRAG 55.7 / 58.4 57.4 / 67.2 68.9 / 79.1 56.1 / 64.7 34.2 / 39.6 14.8 / 13.0 47.9 / 53.7
Structure-Augmented HyperGraphRAG 46.5 / 49.5 34.0 / 43.3 66.1 / 77.6 49.9 / 56.6 36.0 / 45.9 17.1 / 16.1 41.6 / 48.2
Structure-Augmented HippoRAG2 60.1 / 63.8 61.0 / 70.2 66.3 / 78.1 55.9 / 64.6 38.5 / 43.1 8.6 / 5.6 48.4 / 54.2
Multi-Step Self-Ask 57.3 / 60.7 59.6 / 68.8 64.7 / 74.4 52.8 / 58.7 29.3 / 31.2 3.7 / 1.6 44.6 / 49.2
Ours TRA-RAG 65.7 / 66.6 65.1 / 73.4 71.2 / 81.8 70.7 / 78.2 48.8 / 53.4 15.1 / 16.1 56.1 / 61.6

The table below reports multi-hop passage retrieval quality, measured by Recall@5 (Rec@5) and Hit-All@5 (Hit-All):

Method HotpotQA (Rec@5 / Hit-All) 2Wiki (Rec@5 / Hit-All) MuSiQue (Rec@5 / Hit-All) Avg (Rec@5 / Hit-All)
Naive Hybrid 81.9 / 65.7 69.7 / 38.4 52.1 / 16.3 63.9 / 40.1
HippoRAG2 87.3 / 75.9 86.5 / 67.1 62.0 / 31.7 71.6 / 58.2
LinearRAG 88.3 / 76.2 87.4 / 72.3 62.9 / 32.1 72.7 / 60.2
Self-Ask 86.8 / 75.9 89.3 / 76.8 62.8 / 32.0 69.9 / 61.6
TRA-RAG 89.8 / 81.6 97.5 / 94.5 72.2 / 43.5 78.7 / 73.2

Ablation Study

Ablations on 2WikiMultiHopQA assess retriever modalities, reranker components, dialogue context chaining, and evidence context size:

Configuration / Variant F1 LLM-Acc Rec@5 Hit-All@5 Note
TRA-RAG Full Model 70.7 78.2 97.5 94.5 Full pipeline with hybrid retrieval, RRF, and dual reranking
w/ Dense (Cosine) only 70.3 77.6 96.9 94.2 Missing exact lexical keyword matching
w/ Sparse (BM25) only 69.5 77.0 95.7 90.4 Semantic variations and synonyms dropped
w/o Reranker 69.3 76.8 96.5 92.6 Dropping sub-query specific evidence selection
w/o Prev. Context 66.2 73.6 97.1 93.9 High retrieval maintained, but generation fails due to broken reasoning history
w/o QD (Single-step Base) 48.0 51.5 71.2 - Removes decomposition, sequential answering, and reranking

The following table demonstrates answer quality (F1 on 2WikiMultiHopQA) under varying context budgets of retrieved Facts (#Facts) and Passages (#Passages):

#Facts \ #Passages 0 Passages 3 Passages 5 Passages (Default) 10 Passages
0 Facts - 57.3 59.8 66.5
10 Facts 48.5 61.5 67.5 68.6
30 Facts (Default) 58.5 66.2 70.7 70.5
50 Facts 57.2 68.1 70.1 69.3

Key Findings

  • Surpassing the Gold Oracle Baseline: On 2WikiMultiHopQA, TRA-RAG achieves 70.7 F1, outperforming the Oracle baseline (60.5 F1) by 10.2 points. While Oracle provides gold raw text passages, TRA-RAG supplies structured Fact units alongside passages, which distills essential reasoning hops and shields the generator from distractors.
  • Superiority Over Triplets and Hyperedges: In an LLM-as-a-judge comparison across 1,000 passages, the TRA structure won 87.25% of head-to-head evaluations against Triplets (mean score 4.907 vs. 4.198) and 98.55% against Hyperedges (4.971 vs. 3.894), exhibiting decisive advantages in Structural Clarity and Atomic Granularity.
  • Context Saturation and Information Overload: Answer performance peaks at 30 Facts and 5 Passages. Expanding context beyond this threshold (to 50 Facts or 10 Passages) causes performance degradation (dropping from 70.7 to 70.1 and 69.3), underscoring that concise, high-purity evidence selection outperforms indiscriminate long-context injection.
  • Significant Indexing and Runtime Efficiency: During indexing, TRA-RAG requires only a single LLM extraction pass (\(1 \times \text{Tokens}\)), avoiding the multi-turn "gleaning" token overhead incurred by LightRAG (\(2 \times \text{Tokens} + \alpha\)) and HyperGraphRAG (\(3 \times \text{Tokens} + \alpha\)). At inference time on Simple QA, it requires 0.75 seconds per query without any runtime LLM calls, running twice as fast as HippoRAG2 (1.53 s) and over ten times faster than Self-RAG (8.45 s).

Highlights & Insights

  • Entity-Centric Reorganization of Knowledge: By restructuring text around central entity keys rather than flattening them into unanchored pairwise edges, the TRA schema avoids context loss while remaining naturally compatible with nested real-world facts.
  • Pseudo-Natural Language Serialization: Flattening JSON objects into readable template sentences resolves the embedding incompatibility of raw JSON syntax, enabling universal off-the-shelf vector and BM25 retrievers to operate effectively on structured records.
  • Question Decomposition Outperforms Graph Traversal: Empirical results prove that a clean, retrieval-optimized fact space combined with straightforward query decomposition surpasses complex graph random-walk heuristics, demonstrating that optimizing the search target is far more impactful than complicating the search algorithm.

Limitations & Future Work

  • Ambiguity in Long Passages: As document length grows, multiple salient entities emerge within a single text block, creating ambiguity in selecting the primary entity for Title without fine-grained chunking or multi-title extensions.
  • Absence of Global Synthesis Mechanisms: The framework targets entity-level and multi-hop fact retrieval rather than corpus-wide query-focused summarization, where community reports or cluster-level summaries remain necessary.
  • Wikipedia-Centric Evaluation: All benchmark evaluations rely on Wikipedia text; evaluating domain adaptation and schema robustness on specialized corpora (e.g., biomedical reports, financial filings, legal contracts) remains an open challenge.
  • vs. GraphRAG / LightRAG: These frameworks depend on multi-iteration triplet gleaning and graph community summaries, incurring heavy indexing and inference costs. TRA-RAG utilizes single-pass TRA extraction and simple RRF matching, achieving higher recall and lower latency.
  • vs. HippoRAG / HippoRAG2: HippoRAG constructs heterogeneous graphs linking triplets to passages and runs Personalized PageRank (PPR) at runtime. TRA-RAG shows that once search units are structured into atomic Facts, expensive graph-walking algorithms become unnecessary.
  • vs. RAPTOR: RAPTOR relies on recursive tree clustering and text summarization, which inevitably discards granular entity attributes. TRA-RAG preserves precise quantitative values and exact relational qualifiers without lossy abstraction.

Rating

  • Novelty: โญโญโญโญโ˜† [A pragmatic departure from triplet graphs, introducing entity-centric TRA structures and atomic Fact search spaces]
  • Experimental Thoroughness: โญโญโญโญโญ [Extensive evaluations across 6 benchmarks, paired with LLM-as-a-judge structural audits, ablation studies, and scaling analyses]
  • Writing Quality: โญโญโญโญโญ [Clear structural tension, precise methodological descriptions, and honest failure case discussions]
  • Value: โญโญโญโญโญ [Delivers immediate practical utility for production RAG systems by reducing indexing and inference costs while achieving SOTA QA accuracy]