Which content an AI task uses

Key takeaways

• Your files stay on your computer; only the text needed goes to the AI.
• The Docsi server stores no content, only credit usage.
• With your own AI, requests go directly to the chosen provider.
• Local models through Ollama or LM Studio do not leave your computer.
• You see, delete and switch off remembered details in the settings.

Local tools and cloud AI

File conversion, PDF merging and text recognition are local workflows. For an AI task, Docsi AI sends your instruction, required source text and relevant conversation context through the Docsi service to the AI provider.

An optional AI connection of your own uses the selected provider and its terms. A local model server is a different processing route; its operation and capability depend on your setup.

Add only the context you need

1. Select the files and excerpts that belong to the specific task.
2. Remove private information that is unnecessary for the result from a working copy.
3. Review extracted and recognised text. A window image can include other visible details.
4. Check the active AI service and start the task with the context you intend to use.

Consider conversation history and memories

Follow-ups can use earlier tasks and results from the same conversation. In settings, you can review what Docsi may retain about you and remove stored memories. Do not include passwords or access keys as sources or task text.

The path of an AI request

When you start a task with Docsi AI, the app sends a signed request to the Docsi server at docsi-api.impulse-apps.de. It contains your message, up to 16 earlier steps of the conversation and the text excerpts of your sources it needs, with source markers. The server checks your subscription, reserves credits and forwards the request encrypted to the language model.

The answer flows back the same way. The Docsi server stores only how many credits were used – not the content. Processing at the model provider is subject to its terms, which Docsi’s privacy policy refers to.

Connecting your own AI

In the settings you can connect your own provider with your own key instead of Docsi AI. The key is stored in your operating system’s keychain and is never handed back to the interface. Requests then go directly from your device to the provider, not through the Docsi server, and use no Docsi credits.

This processing is subject to the chosen provider’s terms, and the provider bills you for it. The Docsi subscription is still required, because it unlocks the app and its local tools.

Working with local models

If you do not want to pass on any text at all, connect a local model through Ollama or LM Studio. Docsi talks to these programs only through your computer’s fixed local address, without a proxy. For Ollama, Docsi uses a context window of 16,384 tokens.

Quality then depends heavily on the chosen model and your hardware: small models run fast but summarise less accurately and use tools less well. Docsi removes source markers without a matching passage for local models too; check results especially carefully.

What Docsi remembers

Docsi can remember short, lasting details you explicitly tell him, such as your field of study, exam date or preferred citation style. He refuses passwords and similar sensitive data. You see all notes in the settings and can delete them individually there.

With the “Docsi may remember things” switch you turn the feature off entirely. Remembered details are sent as context with AI tasks; older conversations, by contrast, only flow into the same conversation, not into new ones.

Using other people’s data responsibly

If your files contain personal data of others – such as customer lists, applications or minutes with names – the General Data Protection Regulation applies. Its principle of data minimisation means in practice: give an AI task only the parts it really needs, and remove names if the task works without them.

For purely local steps such as converting, merging and text recognition, the data does not leave your computer. For work use, check whether your employer has rules for AI tools, and use a local model when in doubt.

An example: which option fits?

You want to have an application with personal data revised. With Docsi AI the text is processed for the task but not stored; with a local model everything stays on your computer, but quality depends on your model. For a plain conversion to PDF you need no AI at all.

So decide per task: what does the AI really need to see, and is a local step enough? This trade-off is quicker than you might think and protects your data more effectively than any deletion afterwards.

A worked scenario

An example: you want to analyse a customer list with 200 names and addresses to see which regions most orders come from. For this question the AI does not need names. You remove the name column in a working copy and have Docsi analyse only postal code, town and order value.

Docsi performs the calculation itself exactly with his table tool; the AI only phrases the assessment. You get the same result while far less personal data is processed. For working copies that include names, you could alternatively have used a local model.

Key terms explained

Docsi AI

The built-in AI based on the GLM language model by Z.ai, reached through the Docsi server.

Context

All content an AI task sees: the message, earlier steps, source excerpts and remembered details.

Local model

A language model running on your computer, for example through Ollama or LM Studio.

Keychain

The operating system’s protected credential store, where Docsi keeps keys.

Data minimisation

A GDPR principle: process only as much personal data as necessary.