Witivio Documentation

# AI desk Pro AI Models and Terms

AI desk Pro supports multiple AI model providers. The applicable terms and data processing conditions depend on the provider powering the selected model.

# OpenAI models

The default model (GPT-4o mini) is served through Microsoft Azure OpenAI Service. Its use is governed by the Microsoft Product Terms (opens new window) and the Microsoft Data Protection Addendum (DPA) (opens new window).

# Mistral models

Mistral Large 3 is served through Azure AI Foundry. Its use is governed by the Microsoft Product Terms (opens new window) and the Microsoft Data Protection Addendum (DPA) (opens new window).

# Anthropic models

Anthropic models (Claude) are available as Premium models in AI desk Pro and are called directly through Anthropic's API.

Their use is governed by Anthropic's own terms:

Review these terms before you enable Anthropic models on any agent, because data you send to an Anthropic-powered agent is processed under Anthropic's data processing framework.

Selecting an Anthropic model in the agent configuration requires explicit acceptance of Anthropic's terms via a consent checkbox. This step is mandatory and cannot be bypassed.

# Context limits per model

The content AI desk Pro sends to the model on each turn (the system prompt, the extracts retrieved from your knowledge base, and the recent conversation history) is now capped per model instead of sharing a single budget across every model. The cap follows the model that actually serves the request.

This matters for models served through Azure AI Foundry, where the accepted request payload is smaller than the context window advertised for the model. AI desk Pro now trims the retrieved extracts and the history more aggressively for Mistral Large 3, which removes the "payload too large" failures that could make every turn of a conversation fail right after an agent was switched to that model.

Models served through Azure OpenAI (GPT-4o mini) and the Anthropic Claude Premium models keep their previous behavior.

On a long conversation running on a Mistral model, the agent works from fewer retrieved extracts per answer. If an answer looks like it missed part of your content, ask a more focused question or start a new conversation.

← Content Filtering →