SABRE: A Text-to-Database Search Architecture for Relational Databases

Abstract

Relational database queries depend on SQL and prior knowledge of the schema. This work proposes SABRE, a plain-text search architecture for relational databases based on the vectorization of tuples and queries and on retrieval by semantic similarity. The evaluation used a pilot study for selecting the embedding model and experiments with the TPC-H benchmark, comparing the relational and vector modes in terms of retrieval quality, latency, and end-to-end cost. The vector approach achieved a Recall@K of 0.8459, a Precision@K of 0.8039, and an NDCG@K of 0.8886, indicating good coverage of relevant tuples and adequate ranking quality. However, SQL execution showed lower latency in 21 of the 22 queries, and the vector cost was dominated by the database search stage. The results indicate that SABRE is suitable as a complementary mechanism for semantic retrieval and candidate generation, but not as a direct substitute for exact SQL execution in analytical workloads such as TPC-H.