pub const OUTPUT_CEILING_DEFAULT: u64 = 8_192;Expand description
How long one answer may be on the messages face when nobody has said: 8_192 tokens.
The last rung of the output-ceiling ladder on the one face that needs
a figure in every request, reached only when the person stated no
ceiling, the provider’s model list stated none, and the preset table
has no row for the model. The three rungs above it carry real
statements, so this number is never a figure that outranks one
somebody made. The chat and responses faces never reach it: there a
ceiling nobody stated is left to the provider (crates/gateway/Spec.lean §8-17).
Chosen at the width every provider this city calls accepts for every
model it serves, which is what a number used in place of knowledge
has to be. It truncates a long answer rather than refusing the call,
and model_selected records ceiling_from: policy so a truncated
run is read off the account rather than guessed at.