Using the local AI assistant

The local AI assistant lets you ask questions in natural language about the wealth file open in My Wealth for Windows. The model and inference server run directly on your computer: this feature does not send conversation data to a remote AI provider.

1. What the assistant can do

Depending on the installed model and the permissions granted, the assistant can notably:

• summarise your overall wealth position;
• compare assets, investment vehicles, or categories;
• explain a concentration, performance, or allocation;
• find specific information in the file;
• prepare a modification, which remains subject to My Wealth's checks and confirmations.

The assistant is particularly useful for exploring data and preparing a summary. It does not replace My Wealth's calculations or your own verification before a decision or modification.

2. Choosing and installing a model

My Wealth offers three levels to suit your computer's capabilities. The current catalogue offers the following profiles; models and their versions may change with the catalogue displayed in the application.

The target VRAM helps recommend a level. Depending on your configuration, the model may also use system memory and the processor.

1. Open Settings > Local AI.
2. Review the displayed recommendation: My Wealth analyses your computer's capabilities and recommends the most suitable Basic, Medium, or Advanced level.
3. Enable Local AI assistant.
4. Select Add a model.
5. Choose the level suited to your computer and needs.
6. Wait for the download to finish, select the installed model, then choose Apply.

Models are not included in the installer. An Internet connection is required to download or update them; once installed, the model generates its responses locally.

The catalogue presents all available versions compatible with My Wealth that are not already installed. Offers are grouped by Basic, Medium, then Advanced, and sorted by increasing version number within each level. Installing a recent version therefore does not hide older compatible versions still offered. Read each profile’s description, version and hardware requirements before choosing it. A profile’s availability guarantees neither a particular speed nor accurate answers.

Access to the catalogue does not require a Live licence. When a licence is available, it accompanies the catalogue request; it is not a requirement for browsing offers. A list with no new model is distinguished from a connection error.

Available memory and processor or graphics-card performance strongly affect speed. If a model is too slow or does not start, select the level below it.

3. Authorising this wealth file

Permissions are configured separately in each wealth file:

1. open the wealth file you want to query;
2. open Properties > Artificial intelligence;
3. enable read permission for the local assistant;
4. enable modification permission only if you want it to propose writes;
5. apply the properties.

These permissions are stored in the file. They do not automatically apply to another wealth file and remain separate from those granted to MCP clients.

4. Asking a first question

Open the assistant panel, then start with a question whose answer can be checked in the application, for example:

• “What is the total value of my assets and liabilities?”
• “What are my three largest assets?”
• “Where is the highest concentration in my wealth?”
• “What share of my wealth can be accessed quickly?”

Compare the quoted amounts with Situation, Passive income, or Analyses. For a modification request, carefully review the targeted item, proposed values, and confirmation before accepting.

5. Sharing the panel with Checks

Local AI and Checks use the same side panel, with separate tabs. Opening Checks does not delete the conversation. You can check a missing conversion or budget issue, then return to the assistant. Reading issues and using their correction actions does not require AI.

6. Advanced model configuration

By default, My Wealth applies the settings best suited to your computer's capabilities. Automatic mode searches for a compatible backend in the order CUDA, Vulkan, then CPU, while GPU usage is configured to provide good performance without intervention. In most cases, you do not need to change anything.

If you want to favour a specific backend, limit graphics-card use, or troubleshoot a startup problem, you can adjust these settings manually.

Choosing the compute backend manually

CUDA is intended for compatible NVIDIA graphics cards, Vulkan can use other compatible graphics processors, and CPU runs the model without graphics acceleration.

You can manually force CUDA, Vulkan, or CPU. This choice does not silently switch to another backend: if the requested backend does not start, correct its setting or return to Automatic.

Adjusting graphics-card use

This step is optional. Next to the installed model, open Settings, then choose:

• Maximum GPU — recommended to request maximum graphics-card use;
• CPU only to avoid using the graphics card;
• Custom to set the number of GPU layers between 0 and 999.

The number of layers and calibration are saved separately for each backend. A Vulkan setting therefore does not replace the CUDA setting. Run Determine automatically on the backend you want to calibrate.

With Custom, Determine automatically tests several values and displays the current layer. My Wealth then verifies the chosen value with a short local inference. In manual mode, a failure remains associated with the selected backend; in Automatic mode, the test continues with the next backend. If calibration fails, the previous setting is retained: try again, reduce the number of layers, or choose the CPU.

If a selected model is no longer installed, open the settings and choose a model that is still present, or reinstall the corresponding level.

7. Privacy and limitations

8. If you encounter a problem

If the panel remains unavailable:

1. check that an installed model is selected in Settings > Local AI;
2. check read permission in Properties > Artificial intelligence;
3. return to the Automatic backend;
4. try a less demanding model level if memory is insufficient.

9. See also

• Connecting an external application with MCP
• Understanding your wealth analyses
• Solving a common problem