MarketingMind: A Multi-Agent LLM Architecture for Federated Natural Language Querying Across Enterprise Marketing Data Systems
DOI:
https://doi.org/10.15662/IJEETR.2025.0703007Keywords:
multi-agent systems, large language models, federated learning, text-to-SQL, marketing analytics, hallucination mitigation, enterprise data integration, natural language interfaces, CRM intelligence, ReAct framework, OAuth 2.0Abstract
In enterprise marketing environments, there are diverse data ecosystems that encompass customer relationship management (CRM) applications, marketing automation tools, advertising networks, and enterprise data warehouses. One major remaining challenge is the capability to communicate with such systems using natural language interfaces, which is made difficult by the schema heterogeneities, data governance requirements across systems, and the tendency of large language models (LLMs) to hallucinate. MarketingMind is proposed as a multi-agent LLM architecture to handle federated natural language queries in these systems. The framework combines a ReAct-based orchestrator [14] with six dedicated agents responsible for schema discovery, natural language to SQL (NL-to-SQL) translation, hallucination mitigation, federated query routing, CRM analytics augmentation and response synthesis. Retrieval-augmented generation (RAG) [3] is used at various points in the pipeline to anchor the generation of queries to fact-checked schema information. A validation agent that uses self-reflection mechanisms [5] obtains a hallucination rate of 7.3%, whereas the hallucination rate is 38.4% in unconstrained LLM settings. The execution accuracy of the proposed architecture in the Spider benchmark is 91.3%, which is close to the human expert baseline of 94.5%. Federated queries are distributed across six enterprise data sources secured through OAuth 2.0 authentication [2] and column-level encryption, and cover 97.4% personally identifiable information (PII). The architecture is able to handle 138.2 queries per second (QPS) with 500 concurrent users, with a mean latency of 5.4 seconds. MarketingMind highlights that multi-agent orchestration paired with domain-specific RAG grounding and adaptive hallucination mitigation is a viable route towards enterprise-level natural language business intelligence for marketing analytics.
References
[1] R. Basu, N. Aktar, and S. Kumar, "The interplay of artificial intelligence, machine learning and data analytics in digital marketing and promotions: A review and research agenda," J. Marketing Analytics, 2024. https://doi.org/10.1057/s41270-024-00355-6
[2] G. da S. Francisco et al., "Vulnerability detection in intelligent environments authenticated by the OAuth 2.0 protocol over HTTP/HTTPS," Int. J. Comput. Netw. Inf. Secur., vol. 16, no. 2, pp. 1–13, 2024. https://doi.org/10.5815/ijcnis.2024.02.01
[3] Y. Gao et al., "Retrieval-augmented generation for large language models: A survey," arXiv:2312.10997, 2024. https://doi.org/10.48550/arXiv.2312.10997
[4] Z. Hong et al., "Next-generation database interfaces: A survey of LLM-based text-to-SQL," arXiv:2406.08426, 2024. https://doi.org/10.48550/arXiv.2406.08426
[5] Z. Ji et al., "Towards mitigating LLM hallucination via self-reflection," in Findings EMNLP 2023, pp. 1827–1843. https://doi.org/10.18653/v1/2023.findings-emnlp.123
[6] J. Li et al., "Can LLM already serve as a database interface? A Big bench for large-scale database grounded text-to-SQLs," Adv. Neural Inf. Process. Syst., vol. 36, pp. 42330–42357, 2024. https://doi.org/10.48550/arXiv.2305.03111
[7] J. Li et al., "The dawn after the dark: An empirical study on factuality hallucination in large language models," in Proc. ACL, vol. 1, pp. 10879–10899, 2024. https://doi.org/10.18653/v1/2024.acl-long.586
[8] A. Y. Onifade et al., "Advances in CRM-driven marketing intelligence for enhancing conversion rates and lifetime value models," Int. J. Multidiscip. Res. Anal., vol. 4, no. 6, 2024. https://doi.org/10.62225/2583049X.2024.4.6.4326
[9] M. Peeperkorn, T. Kouwenhoven, D. Brown, and A. Jordanous, "The effect of sampling temperature on problem solving in large language models," in Findings EMNLP 2024, pp. 7476–7494. https://doi.org/10.18653/v1/2024.findings-emnlp.432
[10] M. T. H. Rubel et al., "AI-driven big data transformation and personally identifiable information security in financial data: A systematic review," J. Machine Learning, Data Eng. Data Sci., vol. 1, no. 1, pp. 114–128, 2024. https://doi.org/10.58779/jmldeds.v1i01.47
[11] L. Shi et al., "A survey on employing large language models for text-to-SQL tasks," arXiv:2407.15186, 2024. https://doi.org/10.48550/arXiv.2407.15186
[12] W. Switzer et al., "Optimising and evaluating enterprise retrieval-augmented generation (RAG): A content design perspective," in Proc. 2024 8th Int. Conf. Adv. AI, pp. 37–47. https://doi.org/10.1145/3704137.3704181
[13] L. Wang et al., "A survey on large language model-based autonomous agents," Frontiers Comput. Sci., vol. 18, no. 6, p. 186345, 2024. https://doi.org/10.1007/s11704-024-40231-1
[14] S. Yao et al., "ReAct: Synergising reasoning and acting in language models," in Proc. ICLR 2023. https://doi.org/10.48550/arXiv.2210.03629
[15] M. Ye et al., "Heterogeneous federated learning: State-of-the-art and research challenges," ACM Comput. Surv., vol. 56, no. 3, pp. 1–44, 2024. https://doi.org/10.1145/3625558
[16] L. Yu et al., "Mechanistic understanding and mitigation of language model non-factual hallucinations," in Findings EMNLP 2024, pp. 7943–7956. https://doi.org/10.18653/v1/2024.findings-emnlp.466
[17] W. X. Zhao et al., "A survey of large language models," arXiv:2303.18223, 2023. https://doi.org/10.48550/arXiv.2303.18223
[18] X. Zhu et al., "Large language model enhanced text-to-SQL generation: A survey," arXiv:2410.06011, 2024. https://doi.org/10.48550/arXiv.2410.06011
