The Open Data Protocol (OData), developed as a
standard for a RESTful API based on HTTP, is currently being
managed by the Organization for the Advancement of Structured
Information Standards (OASIS) consortium. Querying OData via
natural language has generally gone unexplored. To examine
whether this statement is true or false, I developed Floria, a
web-based SAP Fiori application which uses both WebGPU and
WebLLM/MLC-LLM runtime environments to translate a user’s
natural-sounding question into an OData v4 query. Since all
calculations take place within the browser itself, no part of the
user’s input ever leaves the client. A total of five different-sized
models (from four separate vendors) ranging in size from 0.5
billion to 3.8 billion parameters were tested against a 100-item
benchmark comprised of five different structural query types.
None of these models were pretrained on any OData queries.
Instead, each model was provided with a four-block schema
injection prompt and a six-category post-processor.
In addition, a schema validator was created that would inform
the user of any issues found during validation in plain language
while still allowing them to submit their request. Accuracy per
type is reported for each model.
Index Terms—small language mo

Small Language Models,odata,web gpu,schema constrained inference