🔍 Glossary


An alias for reconstruction.


An object that can convert untrusted input into trusted output.

constraint validator

An object that is capable of analyzing an LLM’s structured output to determine if the output is compliant with a set of constraints.


The process of making a technology, service, or system available to untrusted external users.


A set of rules that define the actual structure of an LLM’s structured output. The grammar is used by a parser to turn some text into a structured tree.


A Large Language Model. A machine learning model that can produce desired text from some prompt. e.g. ChatGPT

prompt envelope

Extra context wrapped around untrusted input to guide an LLM into producing desired output.

prompt injection

A technique for exploiting an LLM by crafting a prompt that causes the LLM to produce output that is considered harmful.


The process of rebuilding LLM’s structured output to be compliant with a constraint validator.


Translating input with a Bifrost.