Definition
Plain language
The piece that reads the model's raw text and spots when it's trying to use a tool.
As stated in the literature
Engine-side logic that extracts structured function calls from generated text, often via regex or partial JSON parsing; a frequent source of injection and memory-safety bugs because it parses untrusted model output.
Also called: tool call parser, tool-call parsing
Why it matters: It is the bridge from a model's words to real actions in the world, and because it parses text the model itself produced, a malformed or manipulated output can push it into behaving in ways nobody intended.
For example, when a model writes out something like a request to look up today's weather, the tool-call parser is what notices, pulls out the city name, and actually calls the weather service.
Heard on the show
“… vLLM's tool-call parser — the part reading the model's raw text, and turning anything that looks like a tool call, into …”Episode 272 — How a Model Guesses Which Engine Is Running It, From a Wrong Date