Generative Business Intelligence in Telecommunications: A Local LLM and RAG-Based Framework for Secure Text-to-SQL Analytics

Traditional business intelligence in telecommunications was built for descriptive, retrospective reporting, relying on batch ETL pipelines, static thresholds, and predefined dashboards. As network complexity, data volume, and customer expectations have grown, these systems have struggled to deliver timely, contextual insight, producing alert fatigue, slow problem discovery, vendor silos, and reactive churn management. Generative Business Intelligence (GenBI), enabled by large language models, retrieval-augmented generation (RAG), and agentic workflows, offers a shift from passive reporting toward active, natural-language analytics. This thesis investigates how Generative BI can be designed and deployed for a telecommunications operator under stringent privacy and on-premise constraints. It develops a seven-phase framework spanning semantic-layer construction, analyst-question example design, time-intelligence handling, model selection, deployment, evaluation, and data governance, and applies it to the text-to-SQL task over a synthetic telecom schema. Two candidate models — a general-purpose cloud API (GPT-3.5) and an open-source, locally deployable specialist (Defog SQLCoder-7B) — are compared on execution success rate, result correctness, runtime, hallucination, and privacy. In this proof-of-concept evaluation, the locally deployed specialist matched or exceeded the cloud model on execution success rate and result correctness, produced no schema-ungrounded queries, ran at comparable speed, and uniquely satisfied the operator's data-locality obligations under the GDPR. The thesis contributes a reproducible methodology for telecom Generative BI deployment that respects data-controller obligations while opening natural-language analytics to non-technical stakeholders.