Questions, answered plainly
Plain answers about where your data goes, who keeps it, which models you can use and how to get access. When we do not know something, we say so.
Where your data goes, and which law applies
What does “EU sovereign” mean here?
Three things. Your account, chats and documents are stored on our servers in Germany. Every model is served by a company established in the EU and processed in the EEA. And this is fixed in our code, so no setting changes it: if a request cannot be served that way, it fails instead of going elsewhere. It does not mean every model was made in Europe; the catalogue says who made each one.
What is eu_native, in plain words?
It is a flag we attach to every request to CORTECS, the EU gateway. It means the company that serves the model must be headquartered in the EU. CORTECS confirmed in writing that this is a hard requirement: if no such company can serve the model, the request fails.
On its own, it does not say where your data is processed. That is data residency: a separate rule, on by default, that keeps your message and the technical data around it in the EEA. Both are true, and neither follows from the other, so we state them separately.
Does my data leave the EU?
Your messages, documents and searches stay in Europe unless you switch on Google web search. Your messages are processed in the EEA, and everything we store is in Germany. Google web search sends the search words for that answer to Google in the United States. You must switch it on and confirm it for each chat. It is not available in Ephemeral chats, and any answer that used it is marked in red. The interface cannot show the green label for that answer.
Who can see my messages?
The company that runs the model sees your message only while it answers, and may not store, log or train on it. CORTECS, the gateway, passes it on without storing it. If the assistant searches the web, Staan sees the search words, not your message. We can technically see what we store, like any hosted service. We do not read it for product, marketing or training purposes. Access is limited to running and supporting the service, and it is logged. No one else: no analytics vendor, no advertising network, no data broker. An Ephemeral chat is never stored, so we have no copy of it to see.
Why do you say “EEA” instead of naming a country?
The EEA (European Economic Area) is the EU plus Iceland, Liechtenstein and Norway. We name a country only for what we run ourselves: our servers, your chats and the code sandbox are in Germany. The providers we use work across the EEA and do not keep processing inside one country. Naming one would be false precision. The one exception is where a model maker is based, such as Mistral AI in France.
Who keeps what, and for how long
What is zero data retention?
The company that runs the model keeps no copy of your message. It may not store it, log it or train on it, and this is written into a data-processing agreement. We request it on every message, and no setting can switch it off. If a model cannot run this way, the request is refused and you are told, so you can pick another model.
It used to be a setting, on by default. We removed the setting because every model we offer can run this way. Switching it off gained nothing; it only let the provider keep your text. The app still shows it on each answer.
Does “zero data retention” mean EurekaGPT keeps nothing either?
No. There are two separate places, and we label them separately everywhere. The AI provider keeps nothing: zero data retention is always on. We do store your chats, on our servers in Germany, so you have a history. Saved chats are deleted automatically 90 days after the last message, or after the period you choose. You can delete any chat sooner.
What is Ephemeral mode?
A chat we never save. An Ephemeral chat is never stored: no messages, no title, no files. Your browser holds the chat while the tab is open. It sends the chat with each message, so the model can follow the conversation. You can attach images: they are sent with your messages and never stored. Documents cannot be attached, because attaching one stores it. Once you close the tab, you cannot return to the chat, search it or export it. Google web search is not available in Ephemeral chats.
Are my chats used to train models?
No. We do not use your chats to train AI. The company that answered is not allowed to either: its contract does not allow it to use your messages to train AI. Your documents are indexed only in our own store, for your own searches. We do not sell your content, share it, or use it to profile you.
How do I delete things?
Delete any chat in the app at any time. Saved chats are also deleted automatically after a period without new messages. You choose 7, 30 or 90 days (the default) or 1 year. Remove a document from your Library and its text and search index go with it. Deleting your account removes your workspace and everything in it. A deleted chat leaves our backups as they rotate, within about a month. For a machine-readable copy of your data (Article 20 GDPR), or anything the app does not cover, write to info@logicance.com.
Where the models come from, and who decides
Why do you offer models made in the US?
Because sovereignty here means where your data is processed and which law applies. Claude, Gemini and GPT are among the strongest models available. They can be served by companies established in the EU, in the EEA, with zero data retention, so leaving them out would only take capability away. We never let you assume they are European: the catalogue says who made each model, and the Trust panel under every answer repeats it.
Why do you offer models made in China, and is that safe?
DeepSeek and GLM are open-weight models: their makers, DeepSeek and Zhipu AI (Z.ai) in China, published the model so anyone can run it. They run on European infrastructure. No user content is sent to China; the weights were trained there.
What is worth weighing is the model itself. Every model reflects the choices of the people who trained it, and any model can be one-sided or evasive on some subjects. We state where each model comes from so you can choose.
How do I know which model answered?
Under every answer is a line with a label and the model’s name. Open it and the Trust panel shows the model and who made it, and the provider that served it and where. It also shows whether a copy was kept, whether the web was searched, and how many tokens (units of text) it used. It comes from what the server recorded for that message, so an answer that used Google web search is always labelled Left Europe.
Can I stop a particular model being used?
New chats start on Efficient unless you or the project choose another model. You choose the model for each chat, and you can give a project a default model for its new chats. Every model in the list runs with zero data retention. There is no way yet to block a model for your whole account. If you need one, tell us what it has to do and we will build it for that need.
Who decides what is in the catalogue?
We do, and the list is short on purpose. A model is added only when it can run with an EU-established provider and zero data retention. We must also be able to say who made it. The list changes as models are added or removed; the current one is in the app.
What you can do with it
What is the Trust panel?
The panel you open from the line under each answer. It is closed by default, so it does not get in the way. It shows whether that answer stayed in Europe, the route it took step by step, and who keeps a copy. One click further, it shows the model and who made it, and the provider that served it and where. It also shows whether a copy was kept, the web search option and the tokens used. It shows only what the server recorded: a fact we do not have is left out.
What is web search, and does it leave the EU?
Three options. Off: the model answers from what it knows and any documents you attached. EU search, the default: our server sends the search to Staan, a European search engine run by European Search Perspective SAS in Paris. It gives the results to the model, which never goes online itself, and lists the sources under the answer. Google: Google’s own search, built into Gemini models, which does leave the EU. It works only with Gemini and asks you to confirm a notice for that chat. It is not available in Ephemeral chats, and it marks the answer in red.
What is the code sandbox, and can it reach the internet?
When a calculation, a data check or a chart is the best answer, the assistant writes Python code. The code runs in a sealed container (the sandbox) on our servers in Germany. It cannot reach the internet: the container has no network route out. It also cannot reach our database, our app or your account.
It runs one job at a time, with limits on memory, processes and time. It holds no passwords or keys of ours. You see the code, the result and any files it made.
Can it produce files?
Yes, real files, in two ways. You can export any reply, or a whole chat, as a PDF, Word, PowerPoint or Excel file. Our server creates the file from the text. And files the sandbox creates come back as downloads attached to the message. The model itself never hosts or links a file, so you never get a download link that leads nowhere.
What happens to documents I upload?
Our server reads them and splits them into passages. Scanned pages are read by Mistral Small, a French model running in the EU, with zero data retention: check names and numbers that matter. The passages are indexed with Mistral Embed, a French model, and stored on our servers in Germany. When a question matches, the relevant passages go to the model, and the answer cites the file and the passage. Remove the document and its text and index go with it.
Who runs EurekaGPT, and how to get access
Who operates EurekaGPT?
LOGICANCE S.L., a company registered in Valencia, Spain (tax ID B93934537). It is the data controller for this website and the provider of the app. Full details are in the imprint. It is a small company.
Where are the servers?
In a Hetzner data centre in Germany: the app, the database, your chats, your documents and their index, and the code sandbox. The model providers we use are established in the EU and process requests in the EEA. We say EEA there, not a country, because that is as precise as we can truthfully be.
Is there a data-processing agreement (DPA) for my company?
Yes. Our DPA covers everything described on this site. It lists our sub-processors (the companies that process data for us), where they are and the safeguards. Ask for it before you sign anything: write to info@eurekagpt.eu.
How do I get access?
By invitation, in small batches. Write to info@eurekagpt.eu with who you are, your organisation and role, and what you want the assistant to do. A person will reply. Your invitation includes 14 days of Pro with no card needed, and 25% more usage for your first 12 paid months if you subscribe in time. If you already have an invite code, create your account in the app and enter it there.