Forbidden terms

Toxicity detection

Partially streams response processing. Response is processed sentence by sentence, as sentences arrive. Note that this can strongly decrease performance, and may yield unexpected results on code or structured output
Make sure to add it and enable it in the toxicity detection section in the connection settings.
Minimum confidence required to mark content as toxic.
When a toxic answer is detected, ask the LLM to retry. Additional instructions will be automatically added.

PII detection

It is recommended to use the dedicated internal code env for PII detection, but another one is configured. You can change the code env used for PII detection in Administration > Settings > LLM Mesh.
PII detection is configured to use the dedicated internal code env, but it has not been installed yet. Contact your admin to install it. Install it in the Administration Settings.
Alternatively, you can change the code env used for PII detection in Administration > Settings > LLM Mesh.
Leave empty for auto-detect
Comma-separated. Valid values are en, fr, de, nl, es, it, ja
Minimum confidence value for detected entities to be returned. Allows for the mitigation of false positives

Prompt Injection detection

Make sure to enable it in the prompt injection detection section in the connection settings.
The detection will be based on the jailbreak output score of the model
Minimum confidence required to mark content as injection.
The prompt sent to the judge to determine whether the following userMessage is a prompt injection. Ensure the prompt guides the LLM to respond with unsafe if it detects an injection or safe if not.
You can use the placeholder {{systemPrompt}} to check if the user message specifically attacks the system prompt of the guarded LLM.

Response Format Checker