[Paper Review] Cost-Efficient RAG for Entity Matching with LLMs: A Blocking-based Exploration

Chuangtao Ma, Zeyu Zhang+3 more|arXiv (Cornell University)|Feb 5, 2026

Data Quality and Management0 citations

TL;DR

This paper introduces CE‑RAG4EM, a blocking-guided, cost-efficient RAG framework for large-scale entity matching that uses batch retrieval and generation to reduce computation while maintaining or improving matching quality.

ABSTRACT

Retrieval-augmented generation (RAG) enhances LLM reasoning in knowledge-intensive tasks, but existing RAG pipelines incur substantial retrieval and generation overhead when applied to large-scale entity matching. To address this limitation, we introduce CE-RAG4EM, a cost-efficient RAG architecture that reduces computation through blocking-based batch retrieval and generation. We also present a unified framework for analyzing and evaluating RAG systems for entity matching, focusing on blocking-aware optimizations and retrieval granularity. Extensive experiments suggest that CE-RAG4EM can achieve comparable or improved matching quality while substantially reducing end-to-end runtime relative to strong baselines. Our analysis further reveals that key configuration parameters introduce an inherent trade-off between performance and overhead, offering practical guidance for designing efficient and scalable RAG systems for entity matching and data integration.

Motivation & Objective

• Motivate and address efficiency and scalability gaps in LLM-based entity matching (EM).
• Propose CE‑RAG4EM, a blocking-guided RAG architecture to enable cost-efficient batch retrieval and generation.
• Provide a unified framework to analyze RAG variants for EM, including blocking strategies and retrieval granularity.
• Evaluate CE‑RAG4EM against strong baselines and across design choices to map performance–overhead trade-offs.

Proposed method

• Introduce blocking-based batch construction to group similar records across source/target tables into blocks.
• Perform block-level batch retrieval of contextual knowledge from a knowledge graph (KG) using Top‑k entities/predicates.
• Execute triple search and graph expansion (BFS and neighborhood expansion) on the retrieved KG context.
• Enrich retrieved identifiers with textual descriptions from the KG and apply ranking-based refinement to filter noise.
• Use tailored prompting strategies for per-query and block-level batch generation to produce EM decisions.

Experimental results

Research questions

• RQ1 Can blocking-based batch retrieval reduce retrieval and generation overhead in RAG for EM without hurting accuracy?
• RQ2 How does retrieval granularity (entity/predicate vs. triples) affect EM performance and efficiency?
• RQ3 What are the trade-offs between block-based batch processing and per-query RAG in EM?
• RQ4 How do knowledge enrichment and refinement influence LLM-based EM decisions?
• RQ5 How do design choices (block size, Top-k, and traversal depth) impact performance and cost?

Key findings

• CE‑RAG4EM can achieve comparable or improved EM quality while substantially reducing end-to-end runtime relative to strong baselines.
• Blocking-based batch retrieval amortizes retrieval costs and enables scalable RAG for EM.
• Retrieval granularity and graph traversal strategy significantly influence trade-offs between accuracy and efficiency.
• Knowledge enrichment and instruction-tuning-based filtering help mitigate noise from retrieved knowledge.
• The framework provides practical guidance on parameter settings (e.g., max block size, Top-k) to balance performance and overhead.
• A unified evaluation framework supports principled comparison of RAG variants across EM tasks and knowledge sources.