Domino IQ Paraphrasing Step Silently Translates Non-English Inputs, Breaking Language-Sensitive System Prompts

ENVIRONMENT:

• HCL Domino 14.5.1 on Windows Server 2022 / Linux

• Domino IQ configured with Remote AI Endpoints

• Reproducible on gpt-5.4-nano, gpt-4.1-mini, and comparable LLMs

DESCRIPTION:
In Domino 14.5.x, the built-in AI paraphrasing security step strips out the original language of user inputs before passing data to the final LLM command.
When a non-English string (such as Czech or German) is processed, the internal hardcoded system prompt "Paraphrase the following text:" causes modern LLMs
to not only paraphrase but silently translate the text into English.

Consequently, language-conditional instructions in the primary command (e.g., "reply in the same language as the original email") fail because
the original language context is lost during the initial sanitization phase.

ROOT CAUSE (via DEBUG_DOMINOIQ_LLMPAYLOAD=1):
1. Domino IQ executes a preprocessing call using the system prompt: "Paraphrase the following text:"
2. The LLM translates the non-English email body to English.
3. The main command (e.g., StdReplyEmail) executes against the English text payload rather than the original input.

PROPOSED ENHANCEMENTS:
1. Provide a mechanism (via dominoiq.nsf or notes.ini) to customize or append to the internal paraphrasing system prompt so administrators can enforce language preservation (e.g., "Paraphrase the following text without translating it. Preserve the original language.").
2. Implement granular control inside the Command document within dominoiq.nsf to allow administrators to disable paraphrasing selectively per command, rather than relying exclusively on the global server-wide DOMIQ_DISABLE_PROMPT_PARAPHRASE=1 setting.