Abstract:
Modern data-intensive applications demand intelligent systems capable of managing heterogeneous, highly interconnected data across multiple specialized storage backends. Yet accessing such systems typically requires expertise in multiple query languages, e.g. SQL, Cypher, MongoDB syntax, limiting accessibility for non-technical users. This paper presents a comprehensive architecture that integrates polyglot persistence, combining document stores, graph databases, key-value caches, and relational data warehouses, with Large Language Models (LLMs) to provide natural language query interfaces. Our implementation compares two Google Gemini model variants,Gemini 3 Pro and Gemini 3 Flash, for translating natural language queries into structured operations across PostgreSQL data warehouses, MongoDB document stores, Neo4j graph databases, and Redis caches. Experimental evaluation across 39 queries spanning six categories reveals a clear accuracy-latency trade-off: Gemini 3 Pro achieves 82.1% fully correct translations with average latency of 12.9-26.5 seconds, while Gemini 3 Flash achieves 76.9% accuracy but with significantly reduced latency of 9.0-11.5 seconds (approximately 1.76x faster). Both models achieve 100% combined accuracy (correct plus partial) with zero incorrect translations. Cross-domain validation comparing Traffic/BI (warehouse-centric) with Social Network (graph-centric) applications demonstrates that translation accuracy improves from 60.0% to 82.1% when moving to structured dimensional schemas, while the architecture adapts effectively across fundamentally different workload patterns. Performance analysis reveals that LLM translation time dominates overall latency (>99%), while database execution remains negligible (<55ms), highlighting opportunities for optimization through caching and prompt engineering. This work contributes a generalizable framework for LLM-powered polyglot persistence systems, comprehensive evaluation methodology for natural language database interfaces, and empirical insights into model selection and domain adaptation trade-offs.
