What is Function Calling?

You give the model a list of functions with names, descriptions and a JSON Schema for their parameters. When a request needs one — what's the weather in Lisbon? — the model doesn't answer from memory; it returns a structured call like get_weather with the city filled in. Your code executes it and passes the result back, and the model continues with real data. The key point is that the model never runs anything itself. It only chooses the function and fills in the arguments; execution stays on your side.

Function calling is the mechanism; tool calling is the pattern built on it. It also has a hard limit: schemas cost tokens, and accuracy degrades as the function list grows, which is why most products ship a handful. Velaris keeps connectors out of the prompt entirely — they're data in a searchable capability mesh the agent queries at runtime, so a run only ever holds the schemas of the few functions that step actually needs, whatever the size of the catalog behind it.