🔍 Glossary

autofixing

An alias for reconstruction.

Bifrost

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.

externalizing

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

grammar

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.

LLM

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.

reconstruction

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

traveral

Translating input with a Bifrost.