User Agreement

The following general conditions apply to the use of Mindbreeze AI.

Operating Models

Four operating models are available for Mindbreeze AI:

• Public Cloud
• Government Cloud
• Private Cloud
• Hyperscaler Cloud

Depending on the operating model, functionality may vary as described.

Volume Packages (pkg)

Volume packages are required to use the AI functionality of Mindbreeze AI described in the scope of services. The volume consumed is determined on the basis of the cumulative calls per year and the number of indexed objects.

A call comprises a request from a user or a service and the corresponding response from the Mindbreeze AI Insight Service called. If a call triggers further sub-calls, these are also counted as independent calls.

Note: Calls that do not include a Mindbreeze AI Insight Service do not count toward volume consumption.

Supported Language Models (LLMs)

Language models that support the “vLLM” or “Hugging Face Text Generation Inference (TGI)” interface can be used.

The specified scope of services is tested with the following models:

• RedHatAI/Mistral-Small-3.2-24B-Instruct-2506-FP8
• openai/whisper-large-v3-turbo

When using other supported language models, functionality and quality may vary.

Pre-trained language models are used in the “Public Cloud” and “Government Cloud” operating models. Fabasoft does not train language models. The selection of the models used is made by Fabasoft.

Supported Embedding Models

Embedding models that are available in ONNX format and optimized for retrieval-augmented generation (RAG) can be used.

The specified scope of services is tested with the following models:

• intfloat/multilingual-e5-large
• sentence-transformers/all-MiniLM-L6-v2
• sentence-transformers/multi-qa-mpnet-base-dot-v1

When using other supported embedding models, functionality and quality may vary.

Pre-trained embedding models are used in the “Public Cloud” and “Government Cloud” operating models. Fabasoft does not train embedding models. The selection of the models used is made by Fabasoft.

Supported Reranking Models

Reranking models in ONNX format can be used to improve relevance for retrieval-augmented generation (RAG).

The specified scope of services is tested with the following models:

• cross-encoder/ms-marco-MiniLM-L12-v2
• Alibaba-NLP/gte-multilingual-reranker-base

When using other supported reranking models, functionality and quality may vary.

Pre-trained reranking models are used in the “Public Cloud” and “Government Cloud” operating models. Fabasoft does not train reranking models. The selection of the models used is made by Fabasoft.

Access Rights

The access rights defined in the Fabasphere are taken into account by Mindbreeze AI.

Dependence on Data and Context

The information, texts, analyses or recommendations generated by the AI are based on underlying data, whereby completeness, timeliness and accuracy cannot be guaranteed. The quality and informative value of the AI output depends largely on the context and the formulation of the input (prompts).

No Guarantee for Accuracy or Suitability

No guarantee is given that the content provided by the AI is correct, complete, up-to-date or suitable for a specific purpose. The use of AI output is at the user's own risk.

Restriction

Since documents of an encrypted Teamroom can only be decrypted at the workstation, AI functionality is not available for these documents.