SEM-RAG: Structure-Preserving Multimodal Graph Compilation and Entropy-Guided Retrieval for Telecommunication Standards

Abstract: Telecommunication standards pose a unique challenge for retrieval systems, where accuracy depends on semantic relevance as well as on preserving the structural logic embedded in the documents, including structured relationships embedded in tables, conditions, and formulas. When these elements are flattened into text, critical dependencies are lost, leading to unreliable retrieval. In this paper, we present SEM-RAG, an end-to-end retrieval framework built around two design choices. First, a layout-aware compiler converts text, tables, and formulas into typed graph primitives. Each table cell is linked to its row headers, column headers, predicates, and source coordinates, while each formula is converted into an operator graph tied to nearby symbol definitions. Second, the compiled graph is compressed with Structural Entropy Minimization (SEM), which avoids LLM-based bottom-up clustering during indexing. A Jensen-Shannon alignment layer and a lightweight query controller serve as supporting retrieval components that map user queries to the right subgraphs, while keeping online cost stable. Experiments on TeleQnA, TSpec-LLM, SPEC5G, and ORAN-Bench-13K show that SEM-RAG improves performance on table-heavy and formula-heavy questions, reaches 94.1\% accuracy on TeleQnA and 93.8\% on ORAN-Bench-13K, and cuts indexing-time token usage by a wide margin relative to standard GraphRAG. These results indicate that structure-preserving compilation is a practical requirement for retrieval over telecom specifications, not merely an optional preprocessing step.

• The paper introduces SEM-RAG, a framework that compiles telecom standards into multimodal graphs to preserve layout and structure.
• It applies structural entropy minimization for efficient hierarchical compression, significantly reducing token usage while maintaining multi-hop reasoning accuracy.
• Empirical results show SEM-RAG outperforms conventional RAG models, achieving up to 94% accuracy and drastically reducing retrieval-induced errors.

SEM-RAG: Structure-Preserving Multimodal Graph Compilation and Entropy-Guided Retrieval for Telecom Standards

Motivation and Problem Context

The increasing complexity of telecommunications standards, characterized by densely interleaved hierarchical clauses, relational tables, and notation-heavy formulas, fundamentally challenges conventional retrieval-augmented generation (RAG) frameworks. Unlike generic natural language corpora, the operational semantics of telecom standards are inherently multimodal and depend strongly on explicit structural logic—row-column bindings in tables, conditional applicability, and operator dependencies in formulas. Standard RAG paradigms, especially those relying on text chunking and semantic similarity, induce "structural amnesia," in which layout-driven dependencies are irrecoverably lost at the ingestion phase. General-purpose LLMs and open-domain RAG systems thus frequently return plausible-sounding but factually unsound or operationally dangerous outputs when tasked with standards-intensive queries.

SEM-RAG Framework: Architecture and Core Innovations

SEM-RAG is an end-to-end system for retrieval from telecom standards, framed around two central innovations:

1. Layout-Aware Multimodal Graph Compilation: A dedicated parser compiles text, tables, and formulas into typed graph primitives, preserving two-dimensional and logical document structure. Table cells are linked explicitly to row headers, column headers, predicates, and provenance metadata. Mathematical expressions are decomposed into operator graphs, with operator nodes, ordered operands, and direct links to nearby symbol definitions. This produces a comprehensive multimodal document graph that is structurally faithful to the source specifications.
2. Structural Entropy Minimization (SEM) for Hierarchical Graph Compression: Instead of conventional LLM-driven hierarchical clustering, hierarchy construction proceeds by minimizing two-dimensional structural entropy over the compiled graph. This approach compresses the graph by exploiting topological regularities and local structural density, producing a practical, update-friendly hierarchy suitable for subgraph-level retrieval. LLM calls are limited to a small set of macro-node summaries at the hierarchy’s top, resulting in sharply reduced indexing-time token consumption.

A Jensen-Shannon divergence-driven alignment layer connects graph and text embedding spaces to support natural language queries, while a lightweight MLP controller dynamically routes queries to direct lookup, localized subgraph expansion, or macro-node retrieval, effectively bounding online latency.

Empirical Evaluation and Quantitative Results

Across four benchmark datasets (TeleQnA, TSpec-LLM, SPEC5G, ORAN-Bench-13K), SEM-RAG demonstrates dominant empirical performance on both aggregate and structure-sensitive metrics:

• End-to-End Accuracy: SEM-RAG achieves 94.1% on TeleQnA and 93.8% on ORAN-Bench-13K, outperforming baselines by significant margins.
• Structure-Sensitive Subsets: On formula-heavy queries, SEM-RAG attains 91.5% accuracy, compared to 51.2% (LightRAG) and 58.7% (Standard GraphRAG); on table-heavy queries, similar advantages are observed.
• Multi-Hop Reasoning: SEM-RAG maintains high performance with increasing reasoning depth, achieving 96.2%, 92.5%, and 87.1% accuracy at 1-hop, 2-hop, and 3-hop question depths, vastly outperforming baselines that rapidly degrade with reasoning chain length.
• Hallucination Mitigation: Manual analysis reveals dramatic reductions in mathematical inconsistencies and row/column binding errors. Row/column logic errors fall from ~22% (Standard GraphRAG) to 2.8% with SEM-RAG.
• Indexing Efficiency: SEM-RAG compresses a million-entity corpus with 0.35M LLM tokens and less than one hour of processing, compared to 110M tokens and 78 hours for LLM-centered GraphRAG pipelines.
• Query-Time Latency: Online latency remains competitive (135 ms), with the alignment and routing overheads marginal relative to graph traversal in baseline systems.
• Ablation Study: Removing structure compilation (table or formula compilation) results in the most severe empirical degradation, underscoring that the dominant empirical gains are due to structure-preserving compilation, not merely improved query routing or alignment.

Theoretical and Practical Implications

The core theoretical contribution is the formulation of executable graph primitives for telecom standards—typed subgraphs representing paragraphs, tables (with predicates), and formulas (with operator structure)—and the demonstration that structural entropy-based, semantics-agnostic hierarchical compression supports scalable, high-fidelity retrieval. The practical implication is that for engineering standards, document ingestion must become a knowledge compilation process rather than text flattening, and that RAG pipelines must shift structure preservation earlier in the workflow.

SEM-RAG’s empirical gains generalize across multiple LLM backbones—including closed and open models (GPT-5.2, Claude 3.5 Sonnet, LLaMA-3-70B, Mistral-Large)—providing absolute performance uplifts of 50 percentage points, indicating model-agnostic improvements are due to evidence structure rather than model-specific induction.

Limitations and Open Problems

SEM-RAG does not provide formal correctness guarantees: upstream parsing errors, OCR artifacts, and highly graphical content (e.g., sequence charts, diagrams) are only weakly represented. The domain scope is limited to text, tables, and formulas, with further work required to extend structure-compiling compilers to more graphical or non-symbolic document artifacts. Cross-domain transfer to other standards-heavy fields (e.g., law, regulation, scientific protocols) is proposed as future work.

Impact and Future Directions

SEM-RAG establishes that in high-stakes technical domains—telecommunications, legal, engineering standards—structure-preserving compilation is a systems requirement for factually grounded retrieval, not a minor preprocessing detail. The work further demonstrates that information-theoretic graph compression can drive practical, LLM-efficient hierarchical indexing without loss of multi-hop reasoning fidelity. The methodology is relevant for widespread adoption as standards corpora evolve in both size and schema complexity.

Future research aims to generalize the compiler to graphical modalities, integrate richer multimodal parsing, and extend entropy-guided retrieval strategies to other domains with dense operational semantics. The framework also raises new questions about symbolic reasoning augmentation, error propagation in multimodal parsing, and optimal trade-offs in retrieval pipeline modularity.