Glossary

Private AI glossary: plain definitions of every term

Published 2026-07-04 · updated 2026-07-04

This glossary defines the vocabulary of private AI companions in plain language, and every definition stands on its own. It covers the cryptography (PBKDF2, AES-GCM, nonces, ciphertext), the sync mechanics (last-writer-wins merges, tombstones, cursors), the voice terms (wake words, endpointing, barge-in), and the terms Violet coined, such as honest-empty states and the trust panel.

How to use this glossary

Each definition below is written to stand alone: you can read one entry without reading the rest and still get a complete, true answer. Where a term describes something Violet does, the definition states what the shipped code actually does, not an aspiration. Violet is pre-launch; nothing here implies live apps or availability.

How the pieces fit together

The private AI companion story is one short chain. Your passphrase runs through a key derivation function (PBKDF2) to produce a content key that never leaves your device. That key seals each record with AES-GCM into a locked envelope: one nonce plus one ciphertext. The server stores envelopes under tenant isolation, reconciles concurrent edits with last-writer-wins merges, records deletions as tombstones, and hands devices incremental changes via a sync cursor. Everything else in this glossary is either a property of that chain, a voice-interface term, or a name Violet coined for an honesty behavior.

Violet's coined terms

Three entries here are Violet's own vocabulary rather than industry standards: honest-empty state, trust panel, and plugged ears. They are defined below as coined terms, each naming a concrete behavior in the product that most software leaves unnamed: empty screens that tell the truth, a privacy panel that reads real state, and a microphone that is verifiably not listening.

Definitions

Private AI companion: A personal AI assistant designed so the provider cannot read what you tell it. It relies on architecture rather than policy: content is encrypted on your device with a key derived from your passphrase, and the server stores only ciphertext it cannot open. The companion part means a resident presence that carries context across your day and devices, rather than a chat window you open and close.
End-to-end encryption (E2E): Encryption in which data is sealed on the sender's device and can be opened only on a device holding the key, so no intermediary, including the service provider, can read it. It differs from encryption in transit (HTTPS), where the server reads everything on arrival, and from encryption at rest, where the provider holds the keys. In Violet's protocol the content key is derived on the client and is never sent to the server.
Locked envelope: Violet's plain-language name for the sealed unit its sync server stores: one nonce plus one ciphertext, both base64 encoded. The server stores and returns the envelope opaquely, along with the merge metadata it needs (a version number, a timestamp, and a deleted flag), and cannot open it. The phrase comes from the brand line: the server sees only locked envelopes.
Content key: The AES-256-GCM key that seals and opens a user's records. In Violet it is derived on the device from the user's passphrase with PBKDF2 and is marked non-extractable, meaning the browser's cryptography API refuses to export it. It never leaves the device and is never stored on or sent to the server.
Key derivation function (KDF): An algorithm that turns a secret such as a passphrase into a cryptographic key of a fixed size. A KDF mixes in a salt so two users with the same passphrase end up with different keys, and a good password-based KDF is deliberately slow to make guessing expensive. PBKDF2, scrypt, and Argon2 are common examples.
PBKDF2: Password-Based Key Derivation Function 2, a standard defined in RFC 8018 that applies a hash function many times over a passphrase and a salt. The iteration count is the cost dial: more iterations make every guess slower for an attacker. Violet derives its content key with PBKDF2-SHA256 at 200,000 iterations over a per-account salt.
AES-GCM: The Advanced Encryption Standard in Galois/Counter Mode, an authenticated encryption mode specified in NIST SP 800-38D. It encrypts data and attaches an authentication tag, so decrypting with the wrong key, or decrypting tampered data, fails outright instead of returning garbage. Violet seals every synced record with AES-GCM under a 256-bit key.
Nonce: A number used once: a value that must never repeat under the same key. In AES-GCM the nonce (also called an IV) makes each encryption unique, and reusing one with the same key seriously weakens the protection. Violet generates a fresh random 12-byte nonce for every sealed record and stores it beside the ciphertext, because a nonce is not a secret.
Ciphertext: The unreadable output of encryption. Without the key, ciphertext carries no usable content; with the key, it decrypts back to exactly the original data. A server that stores only ciphertext can hold, return, or lose your data, but it cannot read it.
Zero-knowledge architecture: A product design in which the provider stores user data it has no ability to decrypt, because the keys exist only on user devices. The phrase is industry shorthand and is unrelated to zero-knowledge proofs, which are a distinct formal cryptographic technique. A practical test: if the provider can recover your content after you lose your passphrase, the architecture is not zero-knowledge.
Local-first software: Software that treats the copy of data on your own device as primary, so it works offline and the cloud is an optional synchronizer rather than the owner of your data. A local-first AI keeps your context usable with no server connection at all. Violet's companion can run against a local core, and with cloud sync off its trust panel truthfully reports that nothing leaves the device.
Hybrid privacy cloud: An architecture that splits work by sensitivity: private content stays sealed on the device, while the cloud handles jobs that need scale, such as syncing ciphertext between devices, account identity, and heavy model inference. The cloud is deliberately blind to stored content because it only ever holds locked envelopes. Violet's commercial design follows this pattern.
Managed inference: Cloud model reasoning provided as part of a service, so users do not have to bring their own model key. The text a user sends for reasoning is necessarily processed as plaintext at that moment, which is a different boundary from encrypted storage, and an honest service says so. Violet's inference proxy checks the caller's entitlement, streams the model's real reply without inventing anything, and never logs message contents.
Wake word: A phrase, usually a product's name, that a device listens for continuously in order to activate. Wake-word systems require the microphone to process audio all the time, which is a real privacy tradeoff even when that processing happens on the device. Violet currently ships no wake word: the microphone is live only while the talk button is held.
Endpointing: Deciding when a speaker has finished talking, so a voice system knows the utterance is complete. Automatic endpointing guesses from pauses and can cut people off; explicit endpointing lets the person signal the end themselves. Violet's push-to-talk design makes endpointing explicit: releasing the button ends the take.
Barge-in: Interrupting a voice assistant while it is speaking, the way you can interrupt a person. Without barge-in, users are forced to wait out a spoken reply they no longer want. In Violet, pressing the mic or simply starting to type stops the spoken reply immediately.
Push-to-talk: A microphone model where audio is captured only while the user actively holds a control, like a walkie-talkie. It is the most conservative voice input design because there is no ambient listening to trust or verify. Violet's companion uses push-to-talk: hold the mic button to speak, release it to send.
Speech-to-text (STT): Converting spoken audio into written words, also called speech recognition. Violet's companion uses the browser's built-in speech recognition, shows exactly the transcript the recognizer returned and nothing more, and plainly says so when a browser has no recognizer instead of pretending to listen.
Text-to-speech (TTS): Generating spoken audio from written text, used for an assistant's spoken replies. Violet's spoken replies use the browser's built-in speech synthesis, are off by default until you turn them on, and speak only replies that really arrived; any new input interrupts them.
Proactive attention: An assistant surfacing something useful without being asked, such as a morning summary or a note that an event starts within the hour. Done honestly, every proactive line must be derived from real data, and silence is the correct output when there is nothing real to say. Violet's moments render only from the actual day's data: no data, no line.
Honest-empty state: Violet's coined term for an empty screen that says truthfully that there is nothing to show, rather than filling the space with placeholders, invented figures, or fake activity. Examples from the product: 'Nothing stored on this device yet.' and 'Nothing scheduled. The day is open.' The rule behind it: real data or an honest empty state, never a fabricated one.
Provenance: Where a piece of information came from, and the trail that proves it. In an AI product, provenance means the assistant can point at the source behind a claim or a memory instead of asserting it from nowhere. Violet's rail is to never claim a memory without a source, and grounded surfaces carry provenance lines such as 'from your real day, nothing invented'.
Consent gate: A control enforced outside the AI model that requires the person's explicit go-ahead before a consequential action runs. Because it sits outside the model, no prompt and no model failure can bypass it. Violet's confirmation gate follows this rule: nothing happens without your say.
Session token: An opaque, high-entropy secret a server issues at sign-in and checks on later requests to know who is calling. Violet's cloud sets it as an HttpOnly, Secure, SameSite=Strict cookie and stores only the SHA-256 hash of the token server-side, so a leaked database does not yield usable sessions.
Tenant isolation: Keeping each account's data strictly separated inside a shared system, so one account can never read another's records. In Violet's cloud, every sync query is scoped by the account ID from a validated session, and requests without one are rejected.
Last-writer-wins (LWW) merge: A rule for reconciling two versions of the same record edited on different devices: the higher version wins, ties fall to the later timestamp, and a remaining tie keeps the incoming record so replays are harmless. It is simple and predictable, at the cost of discarding the losing edit. Violet's sync merges every record this way, and deletes compete like any other write.
Tombstone: A marker that records a deletion so the deletion can propagate to other devices; without one, a synced device would reintroduce the deleted item. In Violet's protocol a tombstone keeps only the merge metadata, and the server nulls out the envelope: no ciphertext remains stored for a deleted record.
Sync cursor: A high-water mark, typically a timestamp, that tells the server to send everything changed since that point, so devices sync incrementally instead of refetching everything. Violet's cursor advances to the newest change in each page, never moves backwards, and jumps forward to the server clock on an empty page so clients do not rescan the same empty window.
Trust panel: Violet's coined name for an in-app surface that shows the privacy state as it actually is, read live from the running system rather than written in advance. Violet's trust panel lists the exact local storage keys present on the device, what the cloud can currently see based on the real configuration flags, a live session probe, and working controls to clear local data and sign out.
Plugged ears: Violet's coined phrase for the state in which she is verifiably not listening. In the current build her ears are plugged by default: there is no wake word and no ambient listening, and the microphone captures audio only while the talk button is held.
Data minimization: The principle of collecting and keeping only the data a feature actually needs, and nothing extra on the theory it might be useful later. Violet's waitlist is an example: joining stores a normalized email address and a short source label, and nothing else.
Threat model: An explicit statement of what a system defends against, from whom, and what is out of scope. End-to-end encryption, for example, defends stored content against a curious or breached server, but not against malware on your own device. Honest products publish the boundary instead of implying protection against everything.
Metadata: Data about data: account identifiers, timestamps, record counts, and sizes, as distinct from the content itself. Even with end-to-end encrypted content, a sync server sees when devices sync, how many records change, and roughly how large they are. An honest product enumerates its metadata plainly instead of implying it sees nothing.
Limits of end-to-end encryption: The boundary of what end-to-end encryption protects. It keeps stored and synced content unreadable to the provider, but it does not hide metadata, it does not protect a device that is itself compromised, it does not cover text sent to a cloud model for reasoning at the moment of processing, and it makes a lost passphrase unrecoverable by design. Knowing the limits is part of trusting the property.
Passphrase strength: How resistant a passphrase is to guessing, which comes mostly from length and unpredictability. A slow key derivation function such as PBKDF2 makes each guess expensive but cannot rescue a passphrase drawn from a short list of common choices. NIST guidance favors long passphrases and screening against known-breached passwords over forced composition rules.

Questions

Why does a glossary matter for choosing an AI companion?

Because the privacy claims that matter are technical, and vague words hide weak architectures. If you know what a content key, a locked envelope, and a key derivation function are, you can ask any vendor precise questions: where is the key derived, what does the server store, and what happens if the passphrase is lost. Precise questions get honest products found faster.

Are these definitions specific to Violet?

Mostly no. Terms like PBKDF2, AES-GCM, nonce, tombstone, and metadata are general and defined generally, with a sentence noting how Violet applies them where that is useful. Three terms are Violet's own coinage and are labeled as such: honest-empty state, trust panel, and plugged ears.