Llm-Augmented Natural Language Query Generation For Nosql Inventory Management

Abstract

Successful inventory management mainly depends on effective querying and analyzing large numbers of database records. Conventional database queries necessitating structured query language (SQL) or NoSQL syntax impose an insurmountable barrier to nontechnical users. This paper proposes a state-of-the-art system based on a Large Language Model (LLM) that converts natural language queries into high-efficiency optimized MongoDB queries. The system uses transfer learning approaches to optimize transformer-based models, thus leveraging Hugging Face's Sentence-Transformers library. The models are optimized to identify optimal embeddings for query understanding.

Furthermore, a selection of vector databases was tested for performance in semantic retrieval, where FAISS demonstrated high performance in retrieval speed for embeddings of high dimensionality. The proposed system integrates few-shot prompting techniques to improve contextual query understanding, utilizing LLaMA 3.3-70B to create robust and adaptive queries. By subjecting the performance of various embedding techniques and retrieval approaches to systematic testing, this paper provides a framework for dynamic query-to-database translation optimization, minimizing execution latency while maximizing retrieval accuracy. The results affirm that integrating state-of-the-art LLMs and optimized vector retrieval greatly enhances query performance, simplifying real-time inventory management for nontechnical users.