Enabling non-technical users to query large-scale databases using natural language is a growing challenge in the field of data analysis. This study addresses the problem of generating correctly executable code for the Elasticsearch database, where standard language models often fail due to complex data structures. The proposed solution is PReFoRCE-ES, an iterative agentic architecture developed using LangChain4j which is a Java framework. This is an adapted version of ReFoRCE for non-relational databases that use context compression, self-refinement and consensus enforcement to validate generated queries prior to execution. The Agent was evaluated on a dataset within the telecommunications domain and benchmarked against a baseline GPT-4o. Results demonstrate that the agent outperforms the baseline on complex reasoning tasks, achieving an Execution Accuracy of 52.5% on reasoning queries and 80% on standard retrieval tasks, compared to the baseline’s 0% .The study concludes that agentic workflows are necessary for ensuring reliability in complex logical operations, though this comes at the expense of increased latency.