Weaver-AISecure Workspace

Weaver-AI is a secure workspace and professional assistant

Account · Masking

Keeping names from model providers

Weaver-AI sends your questions to model providers it does not control. Masking replaces the names you list before the text leaves, and puts them back before you read the answer.

What it does

You give Weaver-AI a list of names. Your company, the clients you work for, your product and project names, the people you discuss. When a message goes to a model provider, each of those names is replaced with a placeholder. The provider answers about the placeholder. Weaver-AI puts your names back before the answer reaches you.

The provider never receives the name. You never see the placeholder.

What it costs

Masking is off by default and it is your choice. It is worth knowing what you are choosing.

A little latencyEvery request is rewritten on the way out and on the way back. The work is small and it is not free.
Occasionally a worse answerThe model is reasoning about a placeholder rather than a name. Most of the time this changes nothing. Sometimes the name carried meaning the placeholder does not.
Questions about the name itselfWeaver-AI cannot answer a question about a name it was not shown. Asking whether a client’s name is spelled correctly will not work while that name is masked.

Where masking reaches, and where it does not

This is the part worth reading carefully. A privacy control that overstates its reach is worse than no control, because somebody relies on it.

Where your text goesWhat masking does
ChatMasked on the way out. Your names are restored in the answer.
Meeting notes and summariesMasked on the way out. The notes come back with the real names in them.
Documents you upload to KnowledgeNot sent to a model provider at all. Weaver-AI indexes them on its own machines, so there is nothing to mask.
Document excerpts used to answer a questionMasked, by the same pass that masks the question.
Your list of names itselfNever sent anywhere. It is stored on your account and used only to do the replacing.

Who can change what

The owner turns it onMasking is a decision for the whole account. Everybody in the account pays the latency, so the switch belongs to whoever answers for the account. On a personal account that is you.
Anybody can add a nameThe person who knows a client’s name is the person who should list it. An owner cannot name every client of every colleague, and a list only the owner can edit is a list that stays short and wrong.
Removing is narrowerAn owner can remove any name. Anybody else can remove the names they added. Taking a name off lowers the protection for everybody in the account, and the person who listed it usually knows why it is there.

Setting it up

  1. Open Account, then Set up masking

    The section is folded away until you open it, because most visits to this page are about something else.
  2. Add the names

    Type several at once, separated by commas. Paste a list from a spreadsheet and the line breaks work too. Choose what they are: a company, a product, a person, a place or an identifier.
  3. Turn the switch on

    Nothing is masked until the switch is on, and nothing is masked that is not on the list. A switch with an empty list costs you the latency and protects nothing, so the page says so.
  4. Save

    Nothing is written to your account until you choose Save. Add a name by mistake, or remove one, and Cancel puts the list back as it was.

How many names you can list

Every listed name is checked against every message you send, so a list is a cost your account pays on each request. The limit depends on the kind of account.

AccountNames
Personal50
Education and research250
Enterprise1000

When Weaver-AI cannot tell whether you mask

Masking settings live in a database. If that database cannot be reached, Weaver-AI does not guess.

If your account has masking on, your message is held and you are asked what to do. You can try again in a moment, send it without masking, or put the message back in the composer and edit it. If you choose to send it without masking, that choice is recorded.

What masking does not change

Your documents are indexed on Weaver-AI’s own machines and are not sent to a model provider to be indexed. That is true whether or not masking is on, and it is why turning masking on does not affect how well search works.

An account that masks cannot use a third-party indexing service, and Weaver-AI refuses rather than quietly indexing your documents elsewhere. The reason is worth stating plainly. Indexing turns text into a list of numbers, and there is no way to put a name back into a list of numbers afterwards. Masking works everywhere else because the text comes back and the names can be restored. That is not true here, so the answer is to keep the work in-house rather than to pretend.