From User Prompt to JPA Query: Comparing Tool Calling, Structured Output & Hybrid Approach

With Java, data can be loaded easily from different sources and displayed in a web UI with little effort. However, as soon as users need to search for specific records, applications often rely on complex filter forms. Modern LLMs offer an alternative: they can extract filter criteria from natural language and turn them into suitable database queries.

Using an example project built with Spring AI and Vaadin, we compare three possible approaches. With Function Calling, the model invokes tool functions in the application code and passes the required parameters. This is easy to understand, but in a simple implementation it quickly reaches its limits when handling complex requests involving negations, operators, or multiple values. Structured Output, on the other hand, directly creates a custom filter object and can represent more complex requests in a compact form. A hybrid approach combines both concepts by using a tool call whose parameters themselves contain a structured filter definition.

All approaches are tested with the same queries against local Ollama models. We compare which approach is more reliable, faster, and more cost-efficient, and examine where each of them reaches its limits.