DECISION PLAYGROUNDREAL VELTO MODEL, IN YOUR BROWSER
Type a request and describe the tools an assistant could use. Velto decides: call a tool, ask for missing information, or no valid tool. This is the real tool-calling model (MiniLM encoder + decision head) running on your device, not a script. Nothing you type is sent anywhere.
Loading the model
Downloading Velto (22 MB). Your browser keeps it, so the next visit is instant...
1. Your request
2. The tools
Required inputs are the pieces of information the tool needs, separated by commas (for example order_id, amount).
Waiting for the model
The decision appears here.
Probabilities
-
These are the model's raw scores, not calibrated chances.
How close each tool is to the request
-
Why this decision
-
Speed (ms per decision)
-
Decision log
All built-in examples: what we expected vs what the model decided
Request
We expected
Model decided
ms
The fine print
It is a research baseline. On a public benchmark (When2Call test) it scores about 0.55 macro-F1. It is right on many easy cases and wrong on others, and the built-in examples show both.
The "missing input" check is word matching. It looks for words from the input name (like amount or temperature) in your sentence, so "$25" or "21 degrees" can be reported as missing. A typed-slot detector (numbers, IDs, dates) is the planned fix, to be tested on human-labelled data first.
Probabilities are raw scores, not calibrated. A 90% here is not a 90% chance of being right.
Speed here is the browser, not the engine. In your browser the model runs in WebAssembly on one thread, so a decision takes several milliseconds. The desktop engine measured p50 3.5 ms and p95 5.3 ms on one laptop.
Same model, small numerical differences. Checked against the Python version on 120 real tool lists: the closest tool matched on all 120 and the decision on 118 (the two differences were near-ties).