Retrieval-Augmented Generation (RAG) enhances large language model (LLM) outputs by grounding responses in external evidence. In enterprise analytics environments, relevant evidence spans both unstructured sources, such as policies and incident reports, and structured systems, including customer records and transactional data. Purely vector-based retrieval provides strong semantic recall but cannot reliably enforce relational constraints such as entity scope or temporal validity. Conversely, SQL-centric retrieval guarantees predicate correctness but lacks robustness to paraphrased natural-language queries. This paper introduces a lightweight hybrid vector–relational integration pattern that unifies a relational data mart with a semantic index through joinable evidence packs: top-k text passages linked to structured entities and filtered using governed predicates. We formalize the system model, define a hybrid scoring formulation that combines dense similarity, sparse lexical matching, and constraint validity, and present an orchestration algorithm that enforces policy tags and prompt-budget limits. A reproducible evaluation framework demonstrates quality–latency and compute–freshness trade-offs across dense-only, SQL-only, and hybrid retrieval modes. By preserving traceability to both semantic excerpts and structured records within a unified governance and observability loop, the proposed approach improves evidence coverage while maintaining constraint correctness and auditability, thereby reducing hallucination risk in enterprise RAG analytics systems.
Avirneni et al. (Sat,) studied this question.