Why don't we have lazy grammar matches?

[ExposedCat]

title

[danbev]

This was actually added recently in #9639 to llama-server. This is available as an option for the /completion endpoint but the list has not been updated yet. The documentation for this might end up looking something like this:

grammar_lazy: Enable/disable lazy grammar. When true the grammar will only be applied for the specified grammar_triggers. Default: false

grammar_triggers: Set of trigger words or tokens used when grammar_lazy is true. Each entry in the list should be in the format { "word": "<word|token>", "at_start": <true|false>} . When one of the words/tokens in this set is generated/predicted, the grammar will be applied (if grammar_lazy is true). Default: []

[ExposedCat]

Yeah I've seen that but this is kinda the same as doing [^trigger] instead of .. Why is there a need for grammar_triggers?