An Integrated ETL–OLAP and Semantic Retrieval Architecture for Hybrid Analysis of Structured and Unstructured Enterprise Data

Extract, Transform, and Load (ETL) and Online Analytical Processing (OLAP) provide robust mechanisms for integrating structured data and supporting multidimensional analysis. Still, they offer limited support for semantic access to unstructured textual content. This study proposes an integrated architecture that combines conventional ETL and data warehousing with semantic retrieval based on dense vector representations and similarity search. The architecture separates structured analytical processing from semantic document retrieval while preserving traceability among source documents, warehouse records, and vector representations. The architecture was examined through an implementation-oriented case study using a large document collection of up to 1.58 million PDF files. The study reports descriptive implementation observations on data ingestion, vector indexing, and query execution in the deployed environment. These observations characterize the operational behavior of the implemented system and are not presented as a controlled performance benchmark between OLAP and semantic retrieval, because the two mechanisms address different classes of information needs. The architecture also provides a query-time foundation for downstream Retrieval-Augmented Generation (RAG); however, this study does not assess generative response quality or LLM orchestration as case-study outcomes. Accordingly, this work contributes primarily through the architectural integration of structured analytical processing and semantic retrieval, preservation of document-to-warehouse-to-vector traceability, and practical implementation of the framework over a large enterprise document repository. The study also identifies limitations related to repeated timing measurements, stage-level profiling, independently annotated relevance judgments, and retrieval-effectiveness assessment. These limitations define future work involving Precision@k, Recall@k, MRR, nDCG, controlled workload benchmarking, and systematic evaluation of RAG-generated responses.