1. Overview
VoltaAI sends your requests to AI models run by third-party providers and returns their answers to you. Around this, we run some AI processing of our own: memory, conversation titles and summaries, web search and a prompt-injection filter. This page explains what each of these does with your content, and the controls you have.
AI providers are described by category and country. Their names are available on request from privacy@voltaservices.co.uk, or under a data processing agreement.
2. Training
We do not train or fine-tune AI models on your content.
Some of the providers we send content to have their own rules:
- AI model router, USA (free models): the router does not train on prompts itself. We ask our AI model router to send requests only to providers that do not use your prompts for training.
- AI model provider, EU (Estonia): does not use data for training without opt-in. Community providers it uses may differ.
- AI model provider, USA: does not state whether it trains on the content it receives.
- Image model provider, EU (Spain), for some image models: does not use prompts to train its own models, but may use them for security and to improve its services.
- Embeddings provider: does not use the data for training.
3. Memory
Memory lets the assistant take your earlier chats into account. It works in the web chat and in our apps. It is not used for requests made with pay-as-you-go API keys.
Memory is on by default. We rely on legitimate interests for it, and you can switch it off at any time (see Switching off and deleting memory).
Depending on how the service is configured, memory works in one of two ways:
- Stored facts: each message you send and each reply that is saved is added to a store of facts about your chats, which is rebuilt after every 10 additions. To build the store, the text of your messages is sent to an AI model provider, which extracts the facts and makes embeddings. The store also keeps the text of those messages. When you send a new message, up to 2,000 characters of relevant facts are added to the request as memory context.
- Search over earlier messages: when you send a new message, we use memory search copies (see below) to look for up to 5 earlier messages that closely match it, and add each one's conversation title and a 200-character preview to the request.
Memory search copies. These are the search data we make from your messages. To make them, the text of your messages is sent to an embeddings provider, through our EU AI model provider, which turns it into numerical representations called embeddings. We store each embedding with a fingerprint of the message text. We keep memory search copies in our database, and in a search index where enabled.
Memory can store anything you write, including sensitive information about you. If you do not want something remembered, delete it from your memories or switch memory off.
4. Titles and summaries
- Titles: we title each new conversation automatically. To do this, we send the first 500 characters of your first message and of the first reply to our AI model provider in the USA or, if that fails, to our AI model provider in the EU.
- Summaries: once a conversation reaches 20 messages, we create a summary of it, and we update the summary after 10 or more new messages. We send each message cut to 300 characters, up to 4,000 characters in total, to the same providers. Summaries are shown to you and included when you export a conversation. They are not added to later prompts.
5. What model providers receive
- Your conversation: the conversation history, trimmed to at most 8,000 characters or 30 messages.
- System context: the date, the persona you chose, the skills you have switched on, and any memory context.
- Image requests: your prompt. Some providers receive the prompt as part of the image's web address.
- No account identifiers: we do not add your user ID or your email address. If an API customer includes a user identifier in a request, it is passed on. The web chat does not send one.
- Web search: when a model searches the web for you, our EU AI model provider writes the search queries from your request. Our own search server, hosted by Hetzner in Germany, sends each query to public web search engines, some of them in the USA. The search engines receive the query text, but not your account details or your IP address. The results go back to the same EU provider, which writes the summary used to answer you. The EU provider sees your request and the search results; the search engines see only the queries.
- Image analysis: when you ask a model to analyse an image, the image goes to our AI model provider in the USA or, if that fails, to our AI model provider in the EU.
- Memory search copies: the embeddings provider receives the text of your messages, as described under Memory.
Our international transfers page sets out where each provider is and the transfer details for each route.
6. How long data is kept
With us. Conversations are kept until you delete them or your account is deleted. Deleting a conversation removes its messages and the memory search copies made from them. Facts that memory has already drawn from it stay until you delete them in your profile.
Memories are kept until you delete them. They are not removed automatically when an account is deleted: stored facts, and memory search copies held in a search index where enabled, stay until you delete them in your profile or ask us to remove them.
Images you generate and documents you upload are deleted after 30 days. Our privacy policy has the full table.
With AI providers.
- AI model provider, EU (Estonia): response caches for up to 30 days, and usage metadata for about 24 months.
- AI model provider, USA: keeps submitted content for abuse prevention, without stating a period, and keeps logs.
- AI model router, USA: does not store prompts unless the account opts in, and keeps metadata about requests. We ask our AI model router to send requests only to providers that do not use your prompts for training.
- Image model provider, EU (Spain): prompts and outputs for the term of its contract; its model hosts keep them for up to 30 days.
- Embeddings provider: abuse monitoring in the region where it processes the data.
7. Prompt-injection filter
Prompt injection is an attempt to make an AI model ignore its instructions. Before we send your conversation history to a model, we leave out any of your messages that match patterns used in such attempts. The model does not see those messages. We do not currently tell you when this happens.
Apart from this filter, we do not run automated content moderation.
8. Human oversight and automated decisions
Answers are generated by AI models and reach you without review by a person.
Our administrators can view conversation content and images through our admin tools.
Some decisions about your account are made automatically, such as plan limits, rate limits, wallet checks and moving your account to the Free plan when a subscription ends. Our privacy policy lists them. For any of them, you can write to privacy@voltaservices.co.uk to ask for information about it, give us your view, ask for a person to review it and contest it.
Our list of models updates automatically from our providers' catalogues, which can add or withdraw models.
9. Labelling of AI output
In the chat, replies from AI models are shown differently from your own messages, so that you can tell them apart.
Generated images do not carry a watermark or a machine-readable mark showing that they were made by AI. If you publish AI output, do not mislead people about how it was made.
10. Limits and accuracy
- AI output can be wrong, incomplete, out of date or biased, and can sound confident when it is wrong. Check anything important before you rely on it.
- AI output is not professional advice.
- Memory can bring in information from earlier chats that is out of date or wrong. You can delete memories that are wrong.
- A model sees only part of a long conversation: at most 8,000 characters or 30 messages, plus the system context.
- The Free plan allows 100 messages and 10 images a day in total. API keys have a rate limit of 60 requests a minute and 5,000 a day by default.
11. Restricted models
Some models are restricted. Restricted models are available only to accounts we have given access. Our terms set out the rules for using them.
12. Switching off and deleting memory
- Switch memory off: turn off the memory switch in your profile. Memory is then not used in your chats from then on, and your new messages are not added to memory.
- See your memories: your profile lists them.
- Delete memories: in your profile, delete them one at a time or all at once.
- Switching off does not delete: memories you already have stay until you delete them.
- Deleting a conversation: deleting a conversation removes its messages and the memory search copies made from them. Facts that memory has already drawn from it stay until you delete them in your profile.
- Help: write to privacy@voltaservices.co.uk.