stainedglass_output_protection

Python package for Stained Glass Output Protection, a library for tokenwise protecting generative AI outputs in vLLM.

Contains both utilities for generating keys and encrypting/decrypting messages, as well as a vLLM plugin for integrating Stained Glass Output Protection directly into vLLM's generation pipeline.

Modules:

Name	Description
decryption	Utilities to decrypt OpenAI-style Completion/Chat Completions responses.
encryption	Utilities for generating keys and encrypting/decrypting messages using the x25519 key exchange and AES-GCM encryption (formatted in the
parsing	Parse decrypted model outputs into reasoning, content, and tool calls.
prompt_embeds_id	Content-addressed identifiers for prompt_embeds payloads.
turboquant	TurboQuant MSE algorithm (arXiv:2504.19874) with encode helpers for the proxy.
vllm	vLLM plugin for Stained Glass Output Protection.