A Hybrid System for Automatic Form Configuration Using LLMs and Deterministic Validation

Form configuration in enterprise low-code platforms such as OmniaPlace involves populating multiple database tables manually, a process that is time-consuming, error-prone, and requires deep platform expertise. This thesis presents a hybrid system that automates this process. Given a SQL DDL schema and a natural language description, the system generates a complete, platform-ready form configuration without manual intervention. The architecture pairs an LLM generation stage with a deterministic five-phase validation pipeline enforcing platform-specific rules. A four-layer prompt architecture encodes domain rules, few-shot examples from real platform exports, and chain-of-thought layout reasoning. A self-healing retry loop feeds structured validation errors back to the model for automatic correction. The system was evaluated across 28 runs on six real form configurations using four prompt versions. An ablation study confirms consistent improvement from v0 to v2. The optimal version (v2) achieves a Field-Type Accuracy of 0.950, a Column Coverage Rate of 0.979, and a First-Pass Validation Rate of 100%. The average Adjusted Time Reduction is 75.2% (83% on the most complex form), with an average specialist edit time of 10.2 minutes. The gpt-5.4-mini variant matches or outperforms gpt-5.4 at one third of the cost. The system is used in production by the eResult operations team.