Skip to main content
One thing, three names. In these docs we call your assistant a Digital Twin. You will also see it called an Instance — on the Instance Settings panel, in the admin Instances list, and in your web address (your-instance.praxislxp.com) — and an Institution, which is what our software calls it behind the scenes (in API fields such as institution). All three mean exactly the same thing: one assistant, with its own knowledge, voice, members and settings. Which name you see depends on the screen you are on, not on what the thing is. An Account is different — it is the organisation that owns one or more Digital Twins.
Can’t find a term? Contact our team at humans@praxis-ai.com to suggest additions to this glossary.

​
Key Definitions

Pria

The flagship Digital Twin created by Praxis AI. Pria serves as both a demonstration of platform capabilities and a practical AI assistant designed to teach prompt engineering skills.

Praxis AI

A human-first AI platform for creating, managing, and using Digital Twins, Assistants, knowledge sources, and integrations. It helps organizations scale expertise, personalize interactions, and support AI-powered workflows.

Digital Twin

An AI assistant configured with its own knowledge, personality, voice, tools, and instructions — a representation of a person, an expert, a brand, or a role. Digital Twins let people interact with expertise in a personalized, scalable way.

Assistant

A reusable set of instructions layered on top of a Digital Twin for one kind of task — a tutor, a grader, a report writer. An assistant adds to the twin; it never replaces it.

Files

Documents or media uploaded to Praxis AI so they can be stored, searched, analyzed, or quoted during conversations. They give the Digital Twin source material for more accurate answers.

Memory

Saved information that helps personalize future interactions, such as your preferences, relevant context, or recurring details. Memory lets Praxis AI give more consistent, context-aware help over time.

RAG

Retrieval-Augmented Generation — Praxis AI searching your uploaded files and answering from what it finds there, instead of only from what the AI learned during training.

Credits

The usage units Pria bills in. Credits let users and administrators see and manage consumption across Digital Twins and services.

Workspace

Your operating environment within Praxis AI, where you access and manage conversations, Digital Twins, files, Assistants, settings, and related resources. (An Agent Workspace is something different — see below.)

Integrations

Connections between Praxis AI and external systems, tools, or services so information and workflows can move between platforms — accessing connected data, automating actions, or extending a Digital Twin’s capabilities.

​
A

The top-level organisation in Praxis AI — a university, a company, a school district. An Account groups one or more Digital Twins, holds the shared credit balance, and lets one set of administrators oversee users across all of them. An Account is not the same thing as an Institution (see I).
A role that gives someone access to manage several Digital Twins within an Account. Account Manager access is configured by the Praxis AI team. Contact humans@praxis-ai.com to request this role.
An admin tool for reproducing user-reported issues. An administrator with the right permissions can temporarily see the product as a user after stating a reason; the session is fully audited, capped at one hour, and the rest of the admin console is locked until the admin clicks Stop. See Memory Parameters & Act As.
The + button at the left of the message box. It opens the menu where you attach files, pick an assistant, choose which knowledge sources and memory Pria may use, and reach other per-conversation settings. Several pages refer to it as “the [+] menu”.
An opt-in mode in which a Digital Twin answers with a team instead of a single model: a manager model (the conductor) plans the work and hires temporary helpers — such as a coder, a tester, or a researcher — that run in parallel and report back. The helpers are separate model instances that share no memory, so each works only from the brief the manager writes for it. Because one request can become dozens of model calls, agent-mode turns cost more than single-model turns. An administrator makes subagents available on a Digital Twin; each member then turns Subagents on for themselves in the [+] menu. The Mode picker (Auto / Basic / Advanced / Expert / Mythical) controls how much team may be hired. See Single Model vs Agent Mode.
A private, isolated cloud desktop that can be switched on for autonomous agent work when enabled for your Digital Twin — its own home directory, configurable CPU, memory, and storage, and an in-browser desktop view. Idle workspaces go to sleep with files preserved and wake on demand.
A capability that extends what a Digital Twin can do beyond writing text — searching Canvas, sending emails, generating images, calling an API, connecting to an external service. Tools are switched on or off per Digital Twin by administrators.
A long secret string (it starts with pria_) that lets your own software sign in to Pria as you, without a browser. It is shown once when created — store it safely. A program exchanges it for a short-lived sign-in token (a JWT) before making calls. See Personal API Keys.
A reusable set of instructions layered on top of a Digital Twin for one kind of task — a tutor, a grader, a report writer. An assistant adds to the twin; it never replaces it: everything the twin already knows — its files, its personality, its connections — still applies. Assistants can be personal (yours), shared with everyone on a Digital Twin, or shared across every Digital Twin in an Account. See Assistants.
Voice memos recorded from the Pria toolbar. Record one or more segments, save, and Pria transcribes them in the background into an Audio Notes collection in your personal vault, replacing the filename with a short AI-generated title. See Audio Notes.

​
B

An enterprise pricing option where organizations supply their own API keys for AI providers (OpenAI, Anthropic, Google, etc.). BYOK customers buy credits at a 40% discount on standard pricing through the Praxis AI sales team, while keeping direct control over provider usage through their existing provider agreements. See Plans and Credits for details.
An advanced enterprise feature that lets organizations connect their own API keys for multiple AI providers — OpenAI, Anthropic, AWS Bedrock, Google (the Gemini Developer API and Google AI Platform / Vertex AI), Mistral, xAI, Moonshot AI (Kimi), and Stability AI — or point Pria at custom endpoints (Azure OpenAI, OpenRouter, RunPod, LiteLLM proxies, self-hosted vLLM). Pria’s Neural Engine then routes each request to the model your configuration selects. See Switch to Praxis-AI Middleware for setup details.

​
C

The connection between Praxis AI and Instructure Canvas LMS. Available via LTI 1.3 protocol or Theme injection, allowing Digital Twins to appear directly within Canvas courses for seamless student and faculty access. Supports two-tier OAuth — a Student Access Token (read-only course data) and a Faculty Access Token (read/write for assignments, announcements, and grades).
An OpenAI-compatible REST endpoint that lets external applications interact with Digital Twins using the standard OpenAI SDK. Point any OpenAI-based app to Pria’s endpoint, use your Digital Twin’s Public ID as the model, and authenticate with a JWT — no other code changes required. See the Chat Completions guide.
One passage of an uploaded file. Pria cuts every file into passages so it can find and quote the right part instead of the whole document. How long the passages are, and how much they overlap, are administrator settings — see Knowledge & RAG Configuration. The docs use “chunk” and “segment” for the same thing.
Pria has two looks. The classic interface is the original one. The new interface, called Studio UI (Beta) on screen, is the redesigned look. An administrator can turn it on for your account (Allow new UI access in Users); inside Canvas it is offered to the Digital Twin’s admins only. Once it is on, the Studio UI button in the top bar switches to it and Classic UI switches back. A few settings live in different menus in each; where the docs give two paths, use the one that matches what you see.
Folder-like groupings for organizing files in your vault, with nesting, drag-and-drop, accent colors, and bulk actions (include/exclude from retrieval, reprocess, download as ZIP). See Organizing Files into Collections.
Pria shortening the older part of a long conversation into a summary while keeping the most recent exchanges word for word, so the twin can remember more without every message costing more. Only what is sent to the AI is shortened — the saved conversation itself is never changed. See Conversation History.
A short check Pria shows when it opens inside Canvas for an admin account: enter the 6-digit code sent to the account’s email address (or from an authenticator app). Students are not asked, and a verified browser is remembered for 24 hours. See More security for Canvas login.
A saved link from a Digital Twin to an external service or an MCP (Model Context Protocol) server, so the twin’s tools can reach that service’s data and actions. See Connectors.
Provenance metadata attached to a generated image or video, following the Coalition for Content Provenance and Authenticity (C2PA) standard, that records which AI model and platform produced it. OpenAI, Amazon, Stability AI and Anthropic attach Content Credentials to the images their models generate; Praxis AI stores generated files as delivered so the credentials stay with the file. See AI-Generated Content Transparency.
A session of back-and-forth dialogue between a user and a Digital Twin. Conversations maintain context and history, can be named for easy reference, and are stored for future retrieval.
Real-time speech-to-speech conversation with a Digital Twin, in many languages: you speak, and the twin speaks back. Powered by a choice of voice providers: OpenAI GPT-Realtime (low-latency, client-side voice selection, tool calling and MCP support), Gemini Live (Google’s native multimodal audio with real-time tool calling and noise reduction), xAI Voice Agent (Grok-powered voice with tool calling and 5 built-in voices), ElevenLabs (custom voice clones, dashboard-managed voices), Cartesia (expressive, tunable voices including clones), OpenAI GPT-Live (listens and speaks at the same time while your Digital Twin’s own model answers), or Anam (a lip-synced video avatar with its own voice catalogue). The voice provider is chosen per Digital Twin in Configuration. Some settings abbreviate it to STS (speech-to-speech). See Convo Mode for usage details.
Praxis Unified Credits — the currency that powers AI interactions on the Praxis platform. Credits are consumed with each request based on the tokens processed (roughly 1 credit per 10,000 tokens, with a 1-credit minimum per call) and never expire. Credits can be pooled at the Digital Twin level or managed individually per user. Models that support caching (such as the Claude and GPT-5 families) can reduce credit consumption by reusing previously processed context — the saving is shown in the Usage Details panel under the reply. See Plans and Credits for pricing and Credit Optimization for saving tips.

​
D

The part of the LTI standard that lets a teacher add Pria to a course from inside the LMS’s own editor — for example as a module item — rather than pasting a link by hand.
One exchange in a conversation: your message and the reply to it. A conversation is a series of dialogues. Settings such as Remember History count dialogues.
A Digital Twin built around a subject rather than a person — a specialist with expert knowledge that is not modelled on anyone in particular. Digital Experts and Digital Twins are created and managed in exactly the same way; only the personality you give them differs. Everything in these docs applies to both, and we say “Digital Twin” throughout.
An AI-powered assistant with its own personality, knowledge base, voice, and appearance. Digital Twins are the core product of Praxis AI — intelligent consultants available 24/7 to answer questions, perform tasks, and provide personalized support. If you registered your own account, you can create a Digital Twin once your personal Pria has answered 5 of your questions (up to 3 per account) — see Create a Digital Twin. You will also see a Digital Twin called an Instance or an Institution — see those entries under I. All three names mean the same thing.
A one-step way to register Pria (or the LXP) with your LMS: instead of typing each setting by hand, the LMS fetches them from a registration link. Available on the LMS platforms that support it.

​
E

Numeric fingerprints of what a passage means. When a document is uploaded to the IP Vault, each passage is turned into one, so Pria can compare your question against all of them at once and find the closest matches by meaning rather than by exact words. This is what powers RAG.
A third-party voice AI provider integrated with Praxis AI as an alternative real-time conversation provider. ElevenLabs supports custom voice clones and branded voice experiences for Digital Twins in Convo Mode. Selected per Digital Twin via the Convo Voice Provider setting.
The process of adding people to a Digital Twin. Users can be enrolled individually, via CSV bulk import, or automatically through Canvas, Google Workspace, or API integrations. When integrated with Canvas or other LMS platforms, users are auto-enrolled on first access. See the Users admin guide for enrollment workflows.
Named permissions that decide what a person may do within a specific Digital Twin — upload files, create assistants, change admin settings, and so on. Each one is written as area.action, for example assistants.edit means “may change assistants” and institutions.edit means “may change this Digital Twin’s settings”. See Entitlements for the full permission reference.
A short-lived pass Pria hands to your browser to start one voice conversation directly with the voice provider. It expires with the call, so no long-lived API key ever reaches the browser.
A premium 1-million-token context window available for select Claude models (Sonnet 5, Sonnet 4.6). The standard context window — how much text a model can take in at once — is 200K tokens. Extended context allows longer conversations and larger document processing but incurs premium pricing from Anthropic. Configured per AI model in Personalization and AI Models.

​
G

Google’s native multimodal audio streaming technology for real-time voice conversations. Supports 30 built-in voices (Puck, Charon, Kore, Fenrir, Aoede, and more), real-time tool calling, noise reduction, and transcription. Uses WebSocket-based streaming with ephemeral token security. Available as a voice provider option in Convo Mode.
A rule about what a Digital Twin must refuse, avoid, or handle with care — written into its instructions.
A simplified version of the interface, with fewer menus, shown to visitors and to newly joined members when an administrator turns it on (Guest UI (Minimum screen) in the Digital Twin’s settings).

​
I

The system that holds your organisation’s logins — Okta, Microsoft Entra ID, Auth0, Google Workspace and similar. With SSO, Pria sends people there to sign in and trusts the answer.
A record created when Praxis Shield flags potentially concerning activity in a Digital Twin’s conversations. Each incident carries a severity, status, and categories, plus a detail report with reviewer notes and an action history. Admins triage incidents — including marking false positives — and nothing is ever enforced automatically.
A file’s status in your vault. Included files are ones your Digital Twin may read and quote. Excluded files are stored but never used in answers. Switch a file either way from the Files panel.
An older name for a Digital Twin. It still appears on some screens — the Instance Settings panel and the admin Instances list — and in your web address (your-instance.praxislxp.com). It means exactly the same thing as Digital Twin. Each Digital Twin has a unique Public ID for integrations.
The name our software uses behind the scenes for a Digital Twin. You will meet it in API fields such as institution, in the admin API reference, and in a few admin screens. It means exactly the same thing as Digital Twin. It is not an organisation or a school in its own right — one organisation typically runs several Digital Twins, each of which is one “institution” in the API. The organisation itself is the Account.
Where you upload files so your Digital Twin can read them. Documents in the IP Vault are cut into passages, indexed, and made searchable, so the twin can answer from your own material and quote it. See IP Vault.

​
J

A short-lived, signed sign-in token that proves who you are on each request. Your browser gets one when you sign in; a program gets one by exchanging an API key. Not to be confused with the tokens that measure text and cost.

​
K

An extra search layer on top of RAG. When a file is processed, Pria can also build a knowledge graph — a map of the people, places, things and ideas the file mentions and how they connect. With KAG on, Pria searches that map as well as the passages themselves and merges the two sets of results into one ranked list (using a method called Reciprocal Rank Fusion). This recovers passages the ordinary search can miss when your wording differs from the document’s. KAG is always an addition to RAG, never a replacement — the ordinary search always runs; there is no “KAG without RAG”. Wherever a knowledge graph is available, Pria uses RAG + KAG Fusion unless you choose RAG, and it remembers that choice.
The settings that control whether, and how, Pria uses your files before answering. There are two: File Vault (formerly Sources) — Disabled (Pria answers without your files), RAG (Pria searches your files by meaning) or, when a knowledge graph is available, RAG + KAG Fusion (Pria also searches the graph and merges the results; used unless you choose RAG). The second setting is the separate Search Only switch, which returns the matching passages as they are instead of an AI-written answer (it has no effect when File Vault is Disabled). Found under Your Profile → Knowledge and in the [+] menu’s Knowledge submenu. See Knowledge Modes.

​
L

The AI that reads your message and writes the reply. Pria can use models from several providers; your administrator chooses which one each Digital Twin uses for each job.
Educational platforms like Canvas, Moodle, Blackboard, and D2L Brightspace. Praxis AI integrates with LMS platforms to provide in-context AI assistance for students and faculty.
The part of Twin Memory a Digital Twin builds on its own: when a conversation changes how it should help you later, it keeps that and shows a Memory row under the answer (Recalled, Saved, Updated, Forgotten). Private to you on that twin, managed in Settings → Memory (new interface) or Manage Memory… → Long-term Memory from the [+] menu’s Knowledge submenu (classic interface), and enabled per twin by its administrators. See Twin Memory.
The industry standard that lets an LMS open an outside tool — such as Pria — with the user and course already identified. Praxis AI supports LTI 1.3 (also called LTI Advantage) for Canvas, Moodle, Blackboard and D2L.
The Praxis Learning Experience Platform — a learner-centric system for delivering personalized, adaptive learning experiences. Features include curated learning paths organized in tracks and lessons, progress tracking, leaderboards, badges, and integrated tools (WordPress, GitLab, Discord). Digital Twins are embedded directly within the LXP to provide AI-powered assistance. See Navigating the LXP.

​
M

An open standard that lets an AI application borrow tools and data from another system. Pria uses MCP in two ways: as a server (offering Digital Twin capabilities to Claude Desktop, Cursor, or any MCP-compatible client) and as a client (connecting to remote MCP servers via Connectors). Connector tools work with every conversation model that supports tools, in typed and voice conversations. See MCP Server for setup.
A Digital Twin sent into a live video call — Zoom, Google Meet, Microsoft Teams or Webex — where it joins as a visible participant with its avatar and voice, listens, answers when addressed, and writes up what happened afterwards. You invite it in conversation (“join <link>, and review the Q3 report”), and the instructions you give in that invitation travel with it into the call. Meetings bill per minute of the bot’s time in the call on top of the usual per-turn cost. See Meeting Agent.
The record that links one person to one Digital Twin, carrying their role (User or Admin), their status, and their entitlements. One person can be a member of several Digital Twins.
Two-step verification (also called two-step sign-in) for Pria accounts. When it is on, signing in from a new browser needs your password (or social sign-in) and a 6-digit code — emailed to the account address (valid 5 minutes) or from an authenticator app. A verified browser becomes a trusted device for 7 days on the same network (24 hours inside Canvas). Admin accounts are asked to turn it on with Strengthen your account security: one that signs in directly must do so, while one that opens Pria from a Canvas course page can select Remind me in 7 days. See Two-Step Verification (user) and MFA Administration (admin).
Information that a Digital Twin retains about users and conversations. There are four systems:
  • Conversation History: Past messages in a specific conversation
  • User Memory: Personal preferences and facts an individual user saves by hand
  • Long-term Memory: What the twin notices and keeps on its own, shown with a receipt in the conversation
  • Shared Memory: Information accessible across all users of a Digital Twin
Long-term Memory and Shared Memory together are called Twin Memory. See Memory Systems.
Pria’s backend — the layer that sits between your organisation and the AI models. It handles sign-in, routes each request to the right model, searches your files, runs tools, and keeps track of conversations, so you don’t have to wire up each AI provider yourself.
A French AI company providing large language models. Pria integrates Mistral via native SDK, supporting conversation (Mistral Large, Medium 3.5, Small 4), code generation (Codestral), text-to-speech (Voxtral TTS), speech-to-text (Voxtral), and embedding models. Models Mistral has retired are replaced automatically by their current equivalent.
The list of AI models Praxis AI supports and keeps up to date, with each model’s abilities and prices. Every Digital Twin picks its models from this shared list. See Platform Models.
One of the jobs a Digital Twin gives to an AI model — conversation, summaries, image generation, embeddings, voice, video analysis, and so on. Each job can be assigned its own model. See AI Models.
Automatic content monitoring that screens user messages and flags inappropriate interactions before they reach the AI model. When enabled, moderation alerts are sent to designated administrators when violations occur. For reviewing and triaging flagged activity after the fact, see Praxis Shield.

​
N

Pria’s model routing layer. The Neural Engine dispatches every AI request — conversation, summaries, image generation, embeddings, voice — to the model your configuration selects, from the assistant’s pinned model down through the Digital Twin’s defaults. With BYOT (Bring Your Own Token) it routes those requests through your own provider credentials and custom endpoints. See Switch to Praxis-AI Middleware.

​
O

Signing in with an account you already have — Google, Microsoft, Apple, GitHub or Facebook — instead of a separate Pria password. Your password is only ever typed at that provider, never at Pria. See Social Sign-In.
Reading the text out of a picture or a scanned page so that it can be searched and quoted like any other document.
The identity layer on top of OAuth 2.0 that lets an organisation’s identity provider tell Pria who just signed in (name, email). This is the standard Pria’s enterprise SSO uses.
A guided questionnaire in the admin panel that helps personalize a Digital Twin. By answering questions about communication style, expertise, and preferences, administrators generate an AI prompt that makes the twin mirror a specific person’s voice and approach. See Onboarding Questions.

​
P

The character a Digital Twin or an assistant plays — its voice, its focus, and how it answers. For a twin it is shaped by its instructions and the Onboarding Questions; for an assistant, by its instructions.
A spot in your LMS where Pria can be opened from — a course navigation link, a module item, and so on.
Two words you can type in a meeting’s chat to steer your Digital Twin. poke asks it for one ordinary spoken answer without a wake word — it waits for the normal turn boundary rather than talking over anyone. poke \<name\> also tells it who is speaking to it, which matters when several people share one screen or profile. hold cancels a pending poke and keeps the twin quiet. See Meeting Agent.
A setting that lets everyone using a Digital Twin draw from one shared credit balance rather than from their own personal credits. When the pool is empty, conversations pause until it is topped up.
Pria’s threat-monitoring capability for Digital Twin administrators. Conversations are scanned in the background; potentially abusive or concerning activity surfaces as incidents in the admin Threats panel, where admins review the report, add notes, progress the status, mark false positives, and share a read-only incident report with another admin of the same Digital Twin. Scanning is invisible to end users and never blocks or alters their experience — admins decide every action. See Praxis Shield.
The flagship Digital Twin created by Praxis AI. Pria serves as both a demonstration of platform capabilities and a practical AI assistant designed to teach prompt engineering skills.
The standing instructions that define a Digital Twin’s (or an assistant’s) personality, behaviour, knowledge focus, and response style. They are handed to the AI at the start of every conversation, which is why they are sometimes called the system prompt. “Prompt” is also used for a message you send to a Digital Twin.
The company whose AI models Pria calls for a given job — OpenAI, Anthropic, Google, Amazon (Bedrock), xAI, Mistral, Moonshot AI, Stability AI for text, images and audio; OpenAI, Google, xAI, ElevenLabs, Cartesia and Anam for voice and avatars.
A short identifier for a Digital Twin, used to point integrations at it — the Web SDK, Canvas, the Chat Completions API. It is public by design: it appears in embed code and in web addresses, so it is safe to share. It is not a password and grants no access on its own; who may use the twin is controlled by sign-in, entitlements, and the twin’s list of authorized URLs. Find it in the Digital Twin’s settings.

​
R

Pria searching your uploaded files before it answers, and grounding the reply in what it finds there instead of only in what the AI learned during training. The search works by meaning, not by exact words, so it finds the right passage even when you phrase things differently from the document. In Pria, RAG searches your personal vault and the Digital Twin’s shared vault. You can add KAG on top, or switch retrieval off entirely — see Knowledge Modes under K. See IP Vault.
An advanced setting that adds a short summary and keywords to each passage before it is indexed, so that searches find it more reliably in long or complex documents.
A setting that controls how long the AI works on a problem privately before it answers. Praxis AI supports five levels: None, Low, Medium, High, and Max. More effort tends to mean better answers to hard questions, but slower and dearer ones. Configured per Digital Twin in Configuration and optionally overridden per AI model. See AI Models for provider-specific details.
How many recent dialogues Pria carries into each new message, so the twin remembers what was just said (0–15; default 3). More history means better continuity and a higher cost per message. See Conversation History.
A person’s permission level within a Digital Twin. Common roles include:
  • User: Standard access to converse and use basic features
  • Admin: Full access to configure settings, manage users, and view analytics

​
S

Having the AI clean junk out of each passage — menus, headers, footers, scanning errors — before it is filed, so searches find the real content.
Ready-made code for putting a Digital Twin into your own app or website. The JavaScript SDK enables web applications to embed Digital Twins with features like messaging, UI control, and event handling.
A Knowledge setting that returns the matching passages from your files as they are, with no AI-written answer on top. Useful when you want to see exactly what your documents say.
A time-bounded period of user activity with a Digital Twin. Sessions track when users are actively engaged and provide analytics on usage patterns.
A type of memory that stores information accessible to all users of a Digital Twin. Useful for institutional knowledge, FAQs, or data that should be consistent across all interactions.
An upload option that stores a file without making it searchable — no processing, no processing cost, and your Digital Twin will not answer from it.
The short, URL-safe name you choose for an SSO connection — my-university in https://your-domain.com/sso/my-university. Lowercase letters, numbers and hyphens only.
Another name for Convo Mode: you speak, and the twin speaks back. Some settings and messages abbreviate it to STS — for example Enabled in Convo (STS) on an assistant.
Signing in to Pria with the account your school or employer already gives you, through your organisation’s identity provider. Set up per Digital Twin by an administrator; each connection gets its own sign-in address. See Enterprise SSO.
Stability AI is a generative-media company focused on image and audio synthesis. Pria integrates Stability AI as a dedicated media-generation provider — powering the generate_image and generate_audio tools available in both chat and convo modes (video generation routes to Amazon Nova Reel or OpenAI Sora instead). Models include Stable Image Ultra / Core, SD 3.5 Large, and Stable Audio 2 (up to 190 seconds). Configured with a platform-provided key or a per-Digital-Twin API key override.
Speech-to-text (STT) turns what you say into written words — dictation, voice transcripts. Text-to-speech (TTS) reads a reply aloud.
The standing, platform-wide guidance Pria sends to the AI with every request, alongside your Digital Twin’s and assistant’s own instructions: how to use tools, which sources to trust first, safety guardrails, formatting rules, and the personality framework. An assistant can be set to bypass it and run on its own instructions alone.

​
T

An imperceptible statistical pattern in an AI model’s word choices that lets the model’s provider recognise text the model generated. It adds no characters or tokens, does not change meaning or quality, and carries no information about the user. Anthropic applies one to Claude output (Fable 5.1 now, Opus 5 from 9 September 2026) and Google applies SynthID to Gemini text, to meet the EU AI Act’s transparency rules; Praxis AI relays replies unchanged so the mark travels with the text. See AI-Generated Content Transparency.
A method of embedding Praxis AI into Canvas by injecting code via the Canvas Theme Editor. This approach provides a branded, always-available Digital Twin experience without requiring LTI configuration.
A small piece of text — roughly four characters, so a little under one word. AI models read and write in tokens, and AI cost and size limits are measured in them. Not to be confused with a sign-in token (JWT) or the ${...} placeholders in assistant instructions (Variables).
Capabilities that Digital Twins can invoke to perform actions beyond text generation. Examples include web search, image generation, code execution, Canvas operations, and memory management.
Where deleted files wait before they are gone for good. You can restore a file from the Trash or delete it permanently from there. Deleting a collection moves its files to the Trash too.
A visual map of your knowledge graph in the Files panel — browse the pillars, entities, and relationships extracted from your documents, down to the source files behind them. Available for any vault that has a knowledge graph: switch the Files display from Files to Tree.
One exchange — a message and its reply. Voice conversations and Agent Mode are billed per turn.
Everything a Digital Twin remembers between conversations: Long-term Memory (what it keeps on its own) plus Shared Memory (what everyone using the twin can draw on). Additional to User Memory. See Twin Memory.

​
U

The panel you open from the Usage Details button under a reply. It shows what that one answer used — tokens, credits, time, and which models ran — including any caching discount. Some pages call it the Dialogue Report Card.
Personal information and preferences that a Digital Twin remembers about an individual user. User memory enables personalized responses that adapt to each user’s communication style and needs over time.

​
V

How Convo Mode works out that you have finished speaking, so it knows when to reply — and, for some voice providers, how readily the twin stops talking when you speak over it. Administrators can tune this per Digital Twin.
Placeholders in assistant instructions, written ${like_this}, that Pria fills in with real values at the moment the assistant runs — for example ${user_name} or ${course_id}. They let one assistant adapt to many situations. See Variables.
Where uploaded files live in Praxis AI. There are three areas: your personal vault (your own files), the Digital Twin’s vault (shared with everyone on that twin — the Instance Files tab in the Files panel), and the Account vault (shared across all Digital Twins in an Account — the Account Files tab). Also called the IP Vault. See IP Vault.
An A–F quality grade for the files in your vault, with categorized follow-ups such as errors, never-used files, stale files, and files that need reprocessing. The Vault Health panel in Files doubles as a filter — click a category to see just those files — and includes a knowledge-graph coverage check. See File Processing Status.
Finding text by meaning rather than by matching words. Vector search is how RAG finds the right passages even when your question uses different terminology from the source material.
The model used to describe the visual content of an uploaded video — the scenes and on-screen content (slides, diagrams, demos, on-screen text). The resulting scene description is fused with the video’s spoken transcript into a single ingested document, so on-screen content becomes searchable in RAG and is captured in the knowledge graph. It is a separate Model Use (independent of the Conversation model), currently limited to video-capable models such as the Gemini Flash and Pro families (gemini-3.8-flash, gemini-3.7-flash, gemini-3.5-flash, gemini-2.5-pro); the platform default is set under System Models and can be overridden per Digital Twin. Runs at low media resolution for cost control and is failure-isolated — a failure never blocks the transcript. See AI Models.
A feature (through the ElevenLabs or Cartesia voice providers) that allows Digital Twins to speak with custom, cloned voices. Organizations can create unique voice identities for their twins.
A feature allowing administrators to audition AI voices before selecting them. Links to OpenAI TTS Playground and Google AI Studio let admins hear voice samples directly.

​
W

The integration method for embedding Digital Twins into custom web applications. The Web SDK provides authentication, UI components, and APIs for full programmatic control.
Two ways your browser can hold a live audio connection open with a voice provider during Convo Mode. Which one is used depends on the provider; WebSocket connections tend to get through strict school and office networks more easily.

​
X

xAI is an AI company founded by Elon Musk. Pria integrates xAI via native API, supporting conversation (Grok 4.7, Grok 4.6, Grok 4.5, Grok 4.3, Grok 4.1 Fast), vision, image generation (Grok Imagine), code generation (Grok Code Fast, Grok Build), text-to-speech (5 voices: Eve, Ara, Rex, Sal, Leo), and real-time voice conversations (Grok-3 Fast via WebSocket). Configured per Digital Twin with a dedicated xAI API key.

​
Z

A guarantee that the AI model provider does not store prompts or model responses after a request completes — nothing is retained on the provider’s servers, used for training, or available for later review. Praxis AI middleware ships only with models that comply with ZDR. The one exception is Anthropic’s Mythos-class models (e.g., Claude Fable 5), which are subject to a mandatory 30-day data retention that overrides ZDR agreements; these models are flagged with a warning in the model selector. See Anthropic’s retention notice.
Can’t find a term? Contact our team to suggest additions to this glossary.