AI models and Ollama

Alcilla does not automatically include one particular language model. You decide which model is used and where it runs. In the typical local scenario, Ollama provides the models.

What is Ollama?

Ollama is a service for running language models on your own workstation or server. Alcilla sends requests to the service and receives the model response.

This may be relevant to you because:

• model processing can take place in your own environment,
• different models can be tested,
• no single cloud provider is imposed technically,
• models can be selected separately for different tasks.

Different models for different tasks

Depending on the configuration, Alcilla can use different profiles, for example for:

• general chat,
• email processing,
• research,
• creation of meeting minutes,
• task extraction,
• titles and short summaries.

A small, fast model can therefore be used for simple tasks while a more capable model handles complex analysis.

Select models without the command line

Alcilla’s model area assists users and administrators with selection:

1. The available hardware is detected.
2. Alcilla evaluates which models are likely to be a good fit.
3. Models can be downloaded directly from the user interface.
4. Approved models then become available in the relevant functions.

If no GPU is detected, many models can still run on the CPU. Processing will generally take longer.

Vision-capable models for image analysis

Image analysis requires a multimodal, vision-capable model. Depending on availability and the Ollama version in use, examples include suitable variants from model families such as Qwen or Gemma.

The vision model is selected separately in Alcilla. This allows a different model to be used for image analysis than for general chat.

Alternative model endpoints

In addition to a simple Ollama setup, Alcilla can—depending on its configuration—connect to multiple model endpoints or compatible providers. This enables scenarios such as:

• a local model for confidential data,
• a central internal model server for multiple users,
• a separate high-performance model for research,
• a combination of different compatible interfaces.

Whether external model providers may be used should always be decided based on protection requirements and internal policies.

Model changes and the knowledge index

If the model used for semantic search is changed, the knowledge index must be rebuilt.

We assist with model selection

Selection is often more difficult than installation. Model size, speed, language, hardware requirements, and answer quality must all fit together.

When needed, we provide support with:

• hardware analysis,
• model comparisons using your sample data,
• Ollama setup,
• selection of chat, search, and vision models,
• deployment of a central model server,
• performance and quality measurements,
• documentation of the selected configuration.