Function-Call Argument Grounding Simulator (2D)

User Utterance

Function Schema

Stats

Slots filled

0/0

Avg match score

0.00

Ambiguous ties

Unfilled required

Grounding Log

Parsing…

How it works

Every function-call argument has to be grounded: a free-text span in the user's sentence has to be matched to one typed parameter of the target function's schema. This engine does that matching for real, in three steps:

1. SPAN EXTRACTION
   quoted strings › ISO/relative dates › numbers › enum-vocabulary words
   › capitalized word sequences (location-like), longest match first,
   non-overlapping.

2. SCORING  (every span × every slot)
   type(span,slot)  = 1.0 exact type match
                       enum: best bigram similarity to slot.values
                       else: small residual cross-type score
   name(span,slot)   = best Dice bigram similarity between the span's
                       neighbouring words and the slot's trigger words
   score = 0.65·type + 0.35·name

3. GREEDY ASSIGNMENT
   sort all (span,slot) pairs by score, descending; walk the list and
   assign the first pair whose span AND slot are both still free.
   A pair is flagged AMBIGUOUS when its winning score is within 0.12
   of that span's (or that slot's) next-best alternative — a real
   near-tie in the computed scores, not a scripted flag.

Change the utterance — pick another example or type your own — and every span, score and assignment is recomputed from scratch. Required slots with no span above the acceptance threshold (0.35) are shown unfilled, exactly like a real function-calling agent that has to ask a follow-up question or fail validation before it can invoke the tool.

Confident match Ambiguous tie Unfilled required slot