enterprise help centers, instructions, account rules, and terms of service are typically organized around articles. However, user questions often involve steps, applicable conditions, and exceptions simultaneously. Under a fixed context budget, traditional retrieval enhancement generation methods that select evidence based on text block relevance tend to have multiple high-scoring blocks from the same article occupying too many positions, leading to the omission of supplementary rule article sources. To address this issue, this paper proposes an article anchor-priority context construction method. This method uses the highest-scoring text block of each candidate article as the article anchor in a unified candidate retrieval result. The system first reserves an evidence position for different articles that meet the eligibility threshold, and then uses the remaining budget to supplement relevant text blocks. This method does not train a new model or calculate pairwise similarity between blocks; it only uses an article eligibility threshold determined by the development set. Based on WixQA enterprise knowledge base data, this paper conducts comparative, hierarchical, parameter sensitivity, and failure tracking analyses under the same candidate pool, token budget, and generation model conditions. Experimental results show that this method mainly reduces the omission of necessary sources and the proportion of single article sources, and achieves more stable and complete evidence coverage in multi-article questions. The response text metric shows only a limited improvement, and the increase in end-to-end latency is minor. The findings demonstrate that the article anchor-first strategy can improve evidence organization and source tracing in enterprise knowledge Q&A with clear and reproducible rules.

Shengping Liu. Design and Implementation of Retrieval Enhancement Generation Method for Enterprise Knowledge Base. International Journal of Business Management and Economics and Trade (2026), Vol. 7, Issue 2: 70-80. https://doi.org/10.38007/IJBMET.2026.070208.

[1] Park SI, Lee JY. Toward Robust RALMs: Revealing the Impact of Imperfect Retrieval on Retrieval-Augmented Language Models[J]. Transactions of the Association for Computational Linguistics, 2024, 12: 1686-1702.

[2] Lee J, Ahn S, Kim D, Kim D. Performance comparison of retrieval-augmented generation and fine-tuned large language models for construction safety management knowledge retrieval[J]. Automation in Construction, 2024, 168: 105846.

[3] Baghdasaryan A, Bunarjyan T, Poghosyan A, Harutyunyan A, El-Zein J. Knowledge retrieval and diagnostics in cloud services with large language models[J]. Expert Systems with Applications, 2024, 255: 124736.

[4] DeBellis M, Dutta N, Gino J, Balaji A. Integrating Ontologies and Large Language Models to Implement Retrieval Augmented Generation[J]. Applied Ontology, 2024, 19(4): 389-407.

[5] Peng Z, Wu X, Wang Q, Fang Y. Soft prompt tuning for augmenting dense retrieval with large language models[J]. Knowledge-Based Systems, 2025, 309: 112758.

[6] Liu, Z. (2026). Research on Measuring User Behavior Response Differences Supported by Propensity Scoring Method. European Journal of AI, Computing & Informatics, 2(2), 47-53.

[7] Zhang, Z. (2026). Research on Performance Monitoring and Data-driven Optimization Mechanisms for AI Systems Oriented Toward Decision Support. Advances in Computer and Communication, 7(2).

[8] Wu, Y. (2026). Research on Interpretable Confidence Rule Base Modeling Method Integrating Data-Driven and Constrained K-Means Optimization. Procedia Computer Science, 279, 612-619.

[9] Wu, Y. (2026). Research on Interpretable Confidence Rule Base Modeling Method Integrating Data-Driven and Constrained K-Means Optimization. Procedia Computer Science, 279, 612-619.

[10] Gao, Y. (2026). Governance Structures and Checks and Balances in Private Equity Funds: Practice, Problems, and Improvement Paths. Financial Economics Insights, 3(2), 82-91.

[11] Wu, Y. (2026). Research on Interpretable Confidence Rule Base Modeling Method Integrating Data-Driven and Constrained K-Means Optimization. Procedia Computer Science, 279, 612-619.

[12] Hou, Y. (2026). Collaborative Regulation for Stable Data Center Operation under Energy Efficiency Constraints. Procedia Computer Science, 281, 612-620.

[13] Zhang, Z. (2026). Research on Performance Optimization Methods for Resource-Aware Model Services in AI Systems. Procedia Computer Science, 281, 1310-1317.

[14] Sun, J. (2026). Automated Feature Engineering and Screening System for Large-Scale Factor Libraries. Procedia Computer Science, 281, 1282-1290.

[15] Chang, C. W. (2026). Privacy Infrastructure Vulnerability Mining and Automated Framework Based on Multimodal AI. Procedia Computer Science, 279, 702-711.

[16] Wu, L. (2026). Construction and Evolutionary Analysis of a Game Model for Supply Chain Finance Funding Based on Blockchain Technology. Procedia Computer Science, 282, 2004-2012.