Security & confidentiality
What happens to your documents, in plain English.
Vorticel is a private AI research assistant for law firms — ask questions across your own documents and get answers grounded in your files, with citations to the exact source passage.
Trusting an AI vendor with privileged client material is not a small ask. This page describes, without marketing language, exactly how your firm's documents are isolated, encrypted, and used — and, just as importantly, exactly what is never done with them. If you're evaluating Vorticel for your firm's IT or compliance review, this page is written for you.
Data flow
How your data flows
Five steps. Here is exactly who touches what, and why it's necessary for search and answers to work — no step skipped, no step hidden.
-
1
Upload
You upload a document over an encrypted connection. It is tagged with your firm's account before anything else happens to it.
-
2
Encrypted at rest
The document is encrypted (libsodium/sodium, per-firm) before it is written to disk. It stays encrypted until your firm requests it back. The original file never leaves this encrypted storage — what leaves in the next step is extracted text only.
-
3
Document text is sent to Voyage AI to be embedded
To make your documents searchable, the extracted text is split into passages and sent to Voyage AI (a MongoDB company) to be converted into numeric embeddings — this is the step that makes retrieval possible, and search cannot work without it. This is the one point where the full text of a document leaves Vorticel's infrastructure. The original file itself never does; only extracted text is sent, and it is sent to embed, not to be read or retained for any other purpose.
-
4
At query time, only the retrieved passages go to Anthropic
When you ask a question, Vorticel searches only your firm's documents and selects a small number of the most relevant passages. Those passages — not the full document, not your document library — are sent to Anthropic (Claude) as ordinary message text to draft the answer. This is never done through a Files API or bulk upload; Anthropic never receives or stores your whole documents, only the specific passages retrieved for that one question.
-
5
No training, ever
Vorticel does not use your firm's documents, questions, or answers to train any AI model, and we send each vendor only the minimum data that step requires — full text to Voyage for embedding, retrieved passages only to Anthropic for answering. We select vendors based on their stated data-handling terms; see the Sub-processors table below for exactly who receives what.
How Vorticel keeps one firm's files separate from another's
Every firm that signs up gets a private account boundary — internally we call it a firm_id. Every document, every extracted passage, every question your team asks, and every answer returned is tagged with that boundary the instant it's created.
Practically, that means: there is no shared search index across firms. When your team asks a question, the system only ever looks inside documents tagged with your firm's boundary — it is structurally incapable of returning another firm's documents, the same way your online banking is structurally incapable of showing you another customer's balance. This isn't a setting that can be misconfigured off; it's how every lookup is written, and it's independently tested before every release.
Within a firm, your team members share access to your firm's documents (role permissions control who can invite teammates or manage billing), but no data ever crosses the boundary between two different firms' accounts.
Encryption at rest
Every document is encrypted before it is stored, using libsodium (the sodium encryption library), with per-firm key handling. Encrypted files are unreadable without the system decrypting them on your firm's authorized request — there is no "plaintext copy on disk" anywhere in normal operation.
Connections between your browser (or the Vorticel sync agent) and our servers are encrypted in transit as well, so documents are never exposed in transit or at rest.
No training on your data
Vorticel does not use your firm's documents, questions, or answers to train any AI model — not for Vorticel's own systems, and not for any third-party model. That is a practice we follow, not a guarantee we are asserting on a vendor's behalf: we select vendors on their stated data-handling terms, and we send each one only the minimum data required for its specific job. Document text goes to Voyage AI to be embedded (see step 3 above); Anthropic never receives more than the specific passages retrieved to answer one question (step 4).
We never use a bulk "Files API" or document-upload integration with Anthropic for the answer-generation step — passages are sent as ordinary message text, scoped to a single question, never the whole document or your library. The original files themselves never leave Vorticel's encrypted storage; what is transmitted is extracted text, and only as much of it as each step actually needs.
Sub-processors
Who receives what, and when.
These are the only outside services that ever receive any of your firm's data, what each one receives, when, and why the product can't work without it.
| Sub-processor | What it receives | When | Why it's necessary |
|---|---|---|---|
| Voyage AI (a MongoDB company) | Text of your documents, chunk by chunk (and the text of each question you ask) | At ingest (documents) and at query time (questions) | To turn text into the numeric embeddings that make search possible — retrieval cannot work without it |
| Anthropic (Claude) | Only the specific passages retrieved for a question, plus that question | At query time, per question | To draft the grounded, cited answer from those passages |
| Stripe | Billing identifiers (your firm, plan, and the email used at checkout) | At billing only | To process subscription payments. Stripe never receives document content. |
| Your Google account identity (email), only if you choose "Continue with Google" | At sign-in only | To authenticate you. Google never receives document content. | |
| Email provider (SMTP) | Recipient address + message text of transactional emails (invites, password resets, notifications) | When such an email is sent | To deliver account email. No document content is ever emailed or attached. |
What is never sent: the original uploaded files never leave encrypted, per-firm storage — only extracted text is embedded, and only retrieved passages (never a whole document or your library) reach Anthropic. No other firm's data is ever present in any payload to any of these services; every request is scoped to your firm's own firm_id.
Common questions
Who else can access my firm's documents?
Your firm's own users, plus two AI vendors that process data as part of answering questions: Voyage AI (a MongoDB company) receives your documents' text to create the search embeddings, and Anthropic receives only the specific passages retrieved to answer a question. No other firm, and no platform staff without a fresh, individually re-verified elevation, can access your documents. See the sub-processors table above for the full list of who receives what.
Does Vorticel send my documents to third parties?
Yes — to make search and answers work, document text is sent to Voyage AI to be embedded, and retrieved passages are sent to Anthropic to draft answers; the original files themselves never leave encrypted storage. Vorticel does not use your documents to train any model, and each vendor receives only the minimum data its step requires. See the sub-processors table above for exactly what each vendor receives and when.
Audit logging
Vorticel keeps an audit trail of security-relevant activity on your account — sign-ins, document uploads and deletions, team membership changes, and administrative actions — scoped to your firm. Platform staff access to your firm's account (for support purposes) is itself logged and requires a fresh, individually re-verified elevation; there is no standing "view any firm" capability, and no impersonation feature exists anywhere in the product.
Data deletion and export
You can delete individual documents at any time; deletion removes both the encrypted file and its extracted, searchable content. You can also download your original documents back out of Vorticel at any time — your firm's files are never locked in.
If your firm closes its account, contact us and we will delete your firm's documents and associated data from active systems, consistent with the retention terms in our Privacy Policy.
How we can prove it
We'd rather you not take the claims on this page on faith. Tenant isolation and grounded refusal are verified, testable behaviors built into how Vorticel ships — not marketing language layered on top. Before any release goes out, an automated check attempts to make one firm's account retrieve another firm's documents and confirms it returns zero rows every time, and a separate check asks Vorticel a question its documents cannot answer and confirms it replies that it couldn't find the answer, without ever calling the underlying model to guess.
We're describing this as verified behavior, not a marketing claim: it is enforced by tests that run before every release, not a promise about intentions. If your firm's IT or compliance reviewer wants to talk through the specifics, see Contact below.
Questions for your IT or compliance team?
We're glad to answer them directly — or start a trial and put the claims on this page to the test with your own documents.