Home▸AI & Machine Learning▸Function-Call Argument Grounding Simulator

Function-Call Argument Grounding Simulator (2D)

A real 2D slot-filling engine: tokenize a user utterance, extract typed entity spans (numbers, dates, quoted strings), score every span against every function-schema parameter with a genuine type-compatibility + string-similarity formula, then greedily assign the best-scoring matches while live-flagging ambiguous ties and unfilled required slots.

AI & Machine Learning2DModerate60 FPS📱 Mobile-adapted⇄ 3D version

Before an LLM agent can invoke a real function, every argument it proposes has to be grounded: a span of the user's own words has to be matched to the right typed parameter of the function's schema. This 2D engine performs that grounding for real — it tokenizes the utterance, extracts typed candidate spans (numbers, dates, quoted strings, enum vocabulary, capitalized names), scores every span against every schema slot with a genuine type-compatibility and string-similarity formula, then greedily assigns the highest-scoring free pairs. Ambiguous near-ties and unfilled required slots are flagged live, and typing a new sentence recomputes the whole assignment from scratch.

⚙ Under the hood

Watch an LLM sample candidate function-call arguments against a JSON-schema valid region in 3D: rejected calls trigger a retry-with-hint loop that nudges the next guess toward validity, with live success-rate and retries-to-completion stats.

function-callingllm-agentsjson-schematool-usetoolformervalidation

2D · HTML5 Canvas 2D · 60 FPS target · runs fully client-side, no install