Natural-language interfaces to sales data typically rely on a large language model to generate executable code, which always returns a number even when it is wrong. This thesis presents an agent that separates query understanding from answer computation. We only call the model twice, once to map the question into a predefined schema built from the data, and once to formulate the results. The model never emits a code or produces a number. Every requested value is calculated by deterministic analyzers, and a separate component verifies all numbers in the generated text. On 105 end-to-end questions the developed system was correct in 86.7 % cases, against 44.8 % for a code-generation baseline
|
