A website can guide a live conversation (SDK): A page that embeds your Digital Twin can now speak into a conversation that is already running — either as the learner, or as the page itself guiding what the twin says next, for example when someone opens a particular lesson. Guiding the twin is off until you allow it: Let the embedding page guide live conversations, in the instance settings beside the transcript sharing switch. Only the page your twin is embedded in can do it, and those turns are billed like any other. The page can also choose the twin’s opening line, or skip the greeting, and is told when a conversation is ready to be spoken into instead of having to guess.Typed messages reach avatar twins: In the classic interface, typing a message during a voice conversation with an avatar twin (Anam or Cartesia) did nothing — the box cleared and the message was never sent. Typed messages now reach the twin on every provider and are treated exactly like speaking them out loud.Importing users never edits an existing account: When you import a list of users (Admin › Users › Import) and an email already belongs to an account, the import only adds that person to your Digital Twin, with the role chosen in the import (a super stays a super). Their account type is raised if the import asks for a higher one and is never lowered. Everything else on the account stays as it is: credits, status, plan, the Digital Twin they currently use, their name, password and MFA. Admin › Users lists people by the Digital Twin they currently use, so an imported existing account appears in your list once the person switches to your twin. To change anything else, open the user in Admin › Users. Through the Admin API,
POST /api/admin/user behaves the same way and lists what it did not apply in ignoredFields; PUT /api/admin/user/{userId} edits an account.Tighter checks on editing accounts (admin): An administrator can edit an account only on a Digital Twin they manage: the one the account currently uses, and any twin they move it to. Only super admins can create, edit, deactivate or delete super admin accounts, and an account type must be user, admin or super.Subagents are now a personal choice (classic UI): An administrator authorizes subagents for a Digital Twin (Instance Settings › Subagents, or the admin editor) and can restrict them to admins with the new Admins only switch. Once authorized, every member sees Subagents in the chat + menu and turns it on for themselves — the choice is saved per twin on your membership. While it is on, a mode control (Auto / Basic / Advanced / Expert / Mythical) sits in the message box next to the Convo button so you can change how much autonomy Pria takes on each request; your pick is remembered. Members who have not turned it on keep single-agent conversations even when the twin allows subagents. The 2026 interface is unchanged.Assistants stay within their Digital Twin: A conversation can only run on an assistant you could pick from the Assistants list — a system assistant, your own personal assistants, the current twin’s assistants, or an account-shared assistant from another twin on the same account. A request that names an assistant outside that set (for example one carried over from a previous twin, or an assistant that has since been deleted) now runs on the standard New Conversation assistant instead of failing or using the other twin’s persona, and switching twins clears the voice assistant picked under the previous one. Earlier conversations that had been recorded against an out-of-scope or deleted assistant are now shown under New Conversation.Subagent turns no longer block the chat (classic UI): when a request is being handled by subagents you can keep chatting. Sending a new message hands the running work to a “Still working” block above the conversation — with its live agent panel and a Stop — and your new message is answered right away. The subagent answer arrives as its own message when it finishes. Reloading the page brings the running work back as the same block.Composer clean-up (classic UI): the red Stop button in the message box is gone — stop a streaming answer with the Abort control under the message, and end a voice session from the Convo controls.Classic sidebar: Settings section removed: The sidebar’s collapsible Settings section duplicated the input + menu, and its collapsed header could show a stray “Instance” line beneath it. Instance and Personal settings live in the + menu (Settings → Instance / Personal); the account Privacy Policy now sits under More.Subagent activity matches the classic transcript: In the classic UI the subagent panel now reads like the other transcript rows — a round agents avatar, a Subagents Details: (N) summary you expand like Tool Details, and the same bordered body — instead of a separate card. The summary also shows the mode the request ran in (Auto, Basic, Advanced, Expert or Mythical) and, on hover, whether it came from your pick in the message box or your saved preference.Subagents Details replays every step: In the classic UI, expanding Subagents Details on a finished reply now shows the steps that produced it, in order — how the request was routed, which model orchestrated it, each subagent it started (with the task it was given and the agent’s card right under it), what it collected and verified along the way, the synthesis, and the model that wrote the answer — instead of only the agent cards.Subagents Details shows where the time went: On replies handled by subagents after this release, the steps replay carries real timings — how long routing, orchestration and the final answer took, how many rounds the orchestrator ran and how much of its time was spent waiting on agents, a Waited line for each pause and which agent it waited for, how long each agent sat in the queue before starting, and the orchestrator’s own narration lines under the Orchestrate step. Earlier conversations keep the step list without the timings.Subagent efficiency (admin): The Charts tab of the Histories admin page gains a Subagent efficiency section for the last 31 days of the current scope — per orchestrator model, mode and surface: turns, median and 90th-percentile duration, where the wall time went (a stacked bar across routing, triage, orchestration and answer), the orchestrator’s rounds, wait ratio, tokens and credits, how the synthesis was produced, and agents per turn with refusals, retries and steers; plus a per-role table of queue, first-tool and duration medians with outcome mixes. The section reports how many subagent turns in the window carried the new timing record. Institution admins see their twin; the same data is available as POST /api/admin/histories/charts/subagent-efficiency (see the Admin API reference).Conversation details expire after 180 days: Your conversations stay complete for as long as you keep them — the question, Pria’s answer, which models answered, and the usage and credits recorded. What is removed after 180 days is the working detail behind an answer: the full text of each tool call and its result, the transcripts of the subagents that worked on it, the passages retrieved from your vault, Pria’s thinking and the memories it recalled. The rows that used to hold those details now say “Details expired after 180 days” and keep their counts (how many passages, tools or thinking rounds there were). Administrators see the cleanup routine’s last runs on the System Health page (System → System Health → Cleanup): the nightly prune and the weekly storage-health pass.System assistants are edited from the admin console: The chat + menu and Manage Assistants no longer offer Edit for a system assistant (one that belongs to no user and no Digital Twin); they are maintained in the admin console, like their deletion already was.One usage breakdown per message: The usage card’s Pricing tab now lists every lane of the reply — the router hop, the orchestrator, the conversation model and each subagent — with tokens in/out, credits, tools and time on one table, and its total shows the credits billed. The separate Subagents tab and the “cheaper than a single premium-model pass” line are gone, and the classic Subagents Details row no longer repeats the receipt.Sign in with Apple: You can now sign in to Pria with your Apple ID — on the iOS app and on the web login page, where Apple sits beside Google, Microsoft, GitHub and Facebook. If you use Apple’s “Hide My Email”, Pria keeps working through Apple’s private relay. If you already have a Pria account under the same email, signing in with Apple simply connects it; you do not get a second account.Dragging a personal file to the Account vault now works: Dropping a file or a collection from My Files onto the Account tab moved the counters as though it had worked, but the file never arrived and the counts snapped back a moment later. Account files live in your instance and are shared across the account, so the move now joins the file to your Digital Twin and shares it in one step. The same fix applies to the bulk Move to Account action and to whole collections, whose files follow the collection.Edited file content appears immediately in the preview: After Edit Content › Save on a text or Markdown file, the preview kept showing the version from before your edit until you closed and reopened it. It now repaints with what you just saved, while the file is re-indexed in the background as before.Connector approvals inside delegated tasks: When Pria hands work to one of her agents and that agent needs a connector tool you have set to require approval, it no longer simply goes without. The agent’s card reads “waiting for your approval” and the request appears under that agent’s activity, in both the classic and the new interface — approve or decline there, or from the pending list in another tab, and the agent carries straight on with the real result. An agent waits up to three minutes. A declined or timed-out call is shown as a failed step rather than a successful one, and the outcome is recorded on the saved message like any other approval. While a plan is building, answer from the waiting agent’s own activity in the build panel; if Pria herself rather than one of her agents needs approval during a plan build, no card appears yet and that step waits out its three minutes without running.Connector approvals stay in sync across tabs and after a reload: A decision on a connector-tool approval card — Approve, Approve for this conversation, or Deny — now updates every open tab, not just the one where you decided: the card used to simply disappear the moment a decision landed, leaving other tabs (and a reloaded page) showing nothing and leaving you unsure whether the tool actually ran. It now shows the recorded outcome in place, with no buttons, wherever the card is open. The tier-upgrade suggestion pill is also now readable in classic dark mode, matching the fix already shipped for the approval card itself.Connector approvals in xAI and Gemini Live voice conversations: The last two voice modes that could not ask now do. When the twin needs approval to run a connector tool, the microphone pauses and the approval card appears — in the transcript, or over the conversation when you are in the compact bubble, the immersive view, or have transcripts turned off. Approve or decline there and the conversation carries on with the real result. In these two modes Pria cannot say the request out loud, so watch for the card; you have 45 seconds. Approve for this conversation is offered when the call is bound to a conversation thread, and if you hang up or reload while a request is waiting, it is cancelled and the tool does not run.Connector approvals now work in voice conversations: In an Anam or Cartesia voice conversation, a connector tool that needs approval no longer just fails to exist. Pria tells you out loud that she needs permission, names the tool and the connector, and puts the same Approve / Deny card in the live transcript — in both the classic and the new interface. Answer, and she carries on with the real result in the same breath. OpenAI Realtime conversations ask too, answered right in your browser (that one is not saved to your approval history, so it cannot be approved from another device). xAI Realtime and Gemini Live conversations do not offer approval-gated tools; their other connector tools now work there for the first time. ElevenLabs conversations will say the approval is not available in that channel and ask you to try again from the Pria chat.Approve a connector tool before it runs: Connectors with Require approval switched on used to hold their tools back entirely — the model was never told they existed. Now the model can ask: a card appears in the conversation naming the tool, the connector and exactly what it would run, with Approve, Approve for this conversation and Deny. Nothing runs until you decide, and the conversation carries on with the real result once you do. The card survives a page reload, a second tab and a second device, and the decision is recorded on the message so you can see later what you allowed. On OpenAI models using their built-in connector support, approval-gated calls used to fail silently — they now ask in the same way. Voice conversations do not show the card yet; there, tools that need approval are simply not offered.MCP connectors work on every model with tool support: A Digital Twin’s connectors (Edit › Connector MCP and Tools) were only used when the conversation model was an OpenAI model; on Claude, Gemini, Mistral or Grok the connector was silently ignored. Pria now connects to the MCP server itself, so the same connectors work on every conversation model, in text chat and in Anam, Cartesia and ElevenLabs voice conversations. Tools run with the same per-server credentials and the same signed identity headers as before. Connectors with Require approval enabled now offer their tools too, gated by the approval prompt described above.New interface no longer demands a course launch from a signed-in user: On phones, and after a Google one-click sign-in inside a Canvas panel, the new interface could show “LMS context required” to someone who was already signed in on a personal account, where the classic interface simply showed the app. It now behaves like classic: that notice appears only when there is no signed-in session at all, and requiring Canvas authentication remains something enforced only on SDK launches from Canvas.Long-term Memory is for everyone once it is on: Switching Long-term Memory on for a Digital Twin, from Instance Settings, the action menu or the admin editor, now also sets Admins only off and the Default Memory Mode to Long-term Memory in the same save, so members get it immediately instead of inheriting the old admins-only, classic-mode defaults. Administrators can still narrow it afterwards.Role Model Pins pick from the model catalog: In Instance Settings › Tools, each subagent role (Planner, Designer, Coder, Tester, Writer, Researcher, Agentic Loop) is now chosen from the same model selector used elsewhere, listing only the models suited to that role, with Default keeping the platform default. No more typing model ids by hand. The Integration wizard’s footer also fits on phones now.Read-aloud voices match the speech model: Choosing Verse or Ballad as the read-aloud voice on a twin using the default tts-1 speech model made every Speak silently fall back to the browser voice, and the Preview button spun and played nothing: those two voices only exist on gpt-4o-mini-tts. The Read-aloud voice picker (Instance Settings, the admin editor and your own Playback voice) now lists only the voices your speech model can speak, adds Fable, Nova and Onyx, and flags a saved voice the model cannot speak as playing as Alloy. Preview shows a clear “Couldn’t play” message on failure, and Speak now tells you when it had to use the browser voice instead of the natural one. The Speech model can now be chosen right there in Instance Settings › Voice, under the natural-voices switch and above the voice, so the two are set together.Join a Digital Twin on your phone: The “Join this Digital Twin” dialog now fills the screen on phones, with the twin’s banner at the top, the About text scrolling in the middle, and Cancel / Join Now pinned at the bottom where your thumb is — instead of floating mid-screen above empty space. On tablets and desktops the dialog is unchanged.Profile avatar aligned in the top bar on phones: The profile picture (or the placeholder icon when you have no picture) at the top right of the classic interface sat slightly below and to the right of its white disc on phones, so the disc and its shadow looked offset. It is now centred.Confirmation before lowering a platform administrator: In Admin › Users, changing a platform administrator’s account type to admin or user now opens a confirmation that spells out what will happen: platform-wide administration is removed and, on save, every one of their memberships is changed too (admin entitlements cleared when changing to user). Cancelling keeps the account as it was.Platform administrators keep their role on every Digital Twin: A platform administrator (super) whose account had been changed to a regular user and then restored kept a plain-user membership on each twin, so the app treated them as a regular user: administration controls disappeared and, on a twin in Guest Mode, they were left with only a Logout button and no way to switch twins. Restoring an account to super now also restores its memberships, the app recognises a super on every twin regardless of the membership row, and a nightly consistency pass repairs any remaining drift.Public twin pages no longer ask visitors for Canvas authentication: A Digital Twin that requires Canvas authentication for its Canvas launches could not be opened through its public web page (/twins/<name>): every visitor was refused with “Access to Canvas LMS requires a valid course Id”. Canvas authentication is now enforced only on Canvas launches; a visitor arriving through the twin’s public page is an anonymous guest with no Canvas course to validate and is let in as before. Canvas launches, and any attempt to use the guest entrance to sign in as a named Canvas user, are gated exactly as they were.Embedded-avatar conversations are now recorded and billed like every other dialogue: Conversations held through a Digital Twin’s embedded avatar widget now appear in the twin’s Histories, attributed to a visitor account created for that visit, with the credits each exchange used. Those credits follow the same rules as any conversation: the twin’s pooled balance, automatic top-ups from the parent account, and any usage limits you have set (per user, per twin, or per 24 hours) all apply to embedded visitors, and the usage reports include them. Visitors who open the widget but never speak are cleaned up automatically. A new Credit Audit view under Admin › System shows, for every twin that pools credits, a daily check that the credits leaving its balance match the credit ledger, with a Run yesterday button to compute a day on demand.Voice transcripts beside the avatar on larger screens: In the classic interface, when a Digital Twin shows transcripts during a voice conversation, tablets, desktops and phones held sideways now place the transcript beside the avatar with the call controls and the message box directly under it, instead of stacking everything in one narrow column. On phones held upright the controls move to the bottom, within thumb reach. The avatar keeps its proportions at every size and with every voice provider. A new Hide transcript button collapses the transcript when you want to focus on the avatar; it comes back expanded on your next conversation. The transcript and the message box are now readable over any background picture, the transcript can be scrolled with the keyboard, and every control is reachable with Tab.Digital Twin sign-in links now work while you’re already signed in: Opening a twin’s sign-in link (a bookmark, an invitation) with a session still active used to leave you on the twin you last used and wrongly warn that you had no access to the one in the link. The link now takes you straight to that twin whenever you’re a member, and the “No access” notice only appears when your account really isn’t allowed in — and says whether the twin is closed to you or simply unavailable. Choosing Log Out & Switch Account from that notice now brings you back to that twin’s sign-in page so you can sign in with the invited address.Live voice transcripts for embedded twins: A page that embeds a Digital Twin through the Pria SDK can now receive the conversation’s transcript live — what the learner says and what the avatar answers, as it happens, for every voice provider — and build its own captions, notes or coaching views from it. It is off by default: the twin’s administrator turns on Share voice transcripts with the embedding page, and the page asks for them with the transcripts display option. Transcripts only ever go to the page hosting the twin.Full-screen avatar appears the moment it connects: Starting a voice conversation directly in the full-screen (immersive) avatar view could leave the avatar’s picture hidden behind the placeholder image even though the call was already live — it only appeared once you started speaking or pressed the mute button. The live avatar now shows the instant its video arrives, so an immersive call looks connected as soon as it is.Avatar greeting timing: The avatar’s introductory message now waits until its video, its audio, the session and the learner’s microphone are all ready (at most five seconds after the first frame), so the opening line no longer lands on top of the connection. A new Greeting delay (ms) setting beside the introductory message (default 1000) controls the pause before it speaks.Long-term Memory everywhere: Long-term memory is now managed the same way in both interfaces. In the classic interface, Manage Memory… (from the action menu’s Knowledge submenu) now opens with a Long-term Memory tab where you can browse, search, correct, forget and export everything your twin has stored — and choose whether it builds memories on its own or sticks to the facts you save by hand — beside the existing User Memory tab. In the new interface, the + menu’s Memory section gains the same one-click Enable Long-term Memory switch platform administrators already had in the classic Knowledge submenu (read-only for Digital Twin administrators). See Twin Memory.A reminder when files are waiting in your Trash: A small notice now appears above the + action menu whenever your personal vault has files in the Trash, so nothing is permanently deleted without you noticing. It tells you how many files are there and how soon the oldest will be purged, and one click opens the Trash so you can restore anything you still need. Administrators with file-edit rights on their Digital Twin see the twin’s Trash count too. Dismiss it and it stays quiet until something changes — a new file lands in the Trash, or a purge draws closer.Change how long a file is kept, right from the preview: The file-info panel beside the preview now shows a Keep if unused setting instead of a read-only date. Files that are still counting down let you pick a longer window or Forever on the spot, with the new “days left” reflected immediately. Files you have opened — previewing counts — or that Pria has used in a conversation are already kept for good, and the panel says so; files in the Trash show their purge date instead.File info on phones no longer squeezes the preview: On phones and small tablets, opening the file-info panel now slides it over the preview as a full-width sheet with its own Close button, instead of shrinking the document to a sliver beside it. The pin and enlarge controls stay where they were on larger screens.Select all in Files → Trash now selects every trashed file: Selecting all in the Trash tab was clearing the selection instead of selecting the files. It now selects every file in the Trash — including the ones not yet loaded — so Restore and Delete Permanently apply to all of them.Files you use are kept for good — and unused ones get 90 days: File longevity now works the way people expected it to. Your personal vault’s default is 90 days (up from 14), and a file only moves to the Trash if it has gone that long without ever being used — using it in a conversation, opening or previewing it, downloading it, or copying a share link all count. The first use is permanent: the file’s countdown ends and its “days left” badge disappears, and it never comes back even if you re-include or restore the file. Digital Twin vaults now keep files forever by default (0 days), so a twin’s knowledge base no longer thins out on its own. Uploading with a countdown shows an amber warning so you can pick a longer window or Forever on the spot. Existing files were brought in line: everyone’s 14-day default became 90, files still on the 14-day window gained the extra time, and every file that had already been used was set to forever — and files the old rule had moved to the Trash were restored.Memory row — see what your twin remembered and what it saved: Whenever an exchange touches your Digital Twin’s long-term memory, the answer now carries a single Memory row — the same expandable style as the Search Results row — listing every memory involved, each tagged with what happened to it: Recalled (brought back to shape the answer, with how it was found), Saved, Updated or Forgotten. Expand a memory to read what Pria remembered; a memory you marked as confidential asks for a second click before it is shown, and one you have since deleted says so. Administrators who prefer a quieter transcript can hide the recalled memories with the new Display Memory Details setting (on by default) in the Digital Twin’s Twin Memory section or in Instance Settings → Knowledge; memories keep working either way, and anything Pria saves, updates or forgets always stays visible. This replaces the earlier one-line memory badge.Turn on long-term memory from Instance Settings: The Knowledge tab of Instance Settings opens with a new Long-term Memory card so administrators can enable the feature and tune it without leaving the product — restrict it to admins while evaluating, choose the default memory mode for new conversations, share memories across the Digital Twins in your account, and control whether recalled memories appear in the Memory row. Enabling the feature itself and account-wide sharing remain platform-administrator actions and show as read-only for other admins. Platform administrators also get a one-click Enable Long-term Memory toggle in the + action menu’s Knowledge → Memory section, whose summary line now reports whether long-term memory is on for the current Digital Twin.Claude Fable 5.1 — Anthropic’s newest flagship model: Claude Fable 5.1 is now available on both the Anthropic API and Amazon Bedrock, for conversation, image analysis, summarization and the agent roles (conductor, planner, coder, tester, writer). It keeps Fable 5’s 1M-token context window and 128K-token responses, and its adaptive reasoning is always on — the model decides how much to think, and the Reasoning Effort setting steers depth from low to max. Prompt-cache reads are billed at 2.5% of the input price (versus 10% on other Claude models), so long, stable system prompts get markedly cheaper across a conversation. Like Fable 5 it carries the 30-day retention marker in the model selector. Select it in Admin → AI Models (claude-fable-5-1 on Anthropic, global.anthropic.claude-fable-5-1 on Bedrock — the Bedrock variant requires the AWS account data-sharing opt-in plus a model-access agreement for this version; see the Platform Models guide).Prevent Personal Account Creation: A new instance setting lets administrators stop members from creating a personal Pria account for themselves. When it is on, the offer no longer appears for anyone working in that instance. Find it in your Digital Twin’s settings beside Prevent Switching to Other Instances. See Configuration.Jump straight to any setting in the admin panels: Editing a Digital Twin, account, or user on a wide screen now shows a quick-access list down the right of each tab. It names every section in that tab, highlights the one you are looking at as you scroll, and jumps to any of them in one click — so finding a setting no longer means scrolling a long form end to end. It appears only where there is more than one section to list, and only on larger screens.Guest twin links are easier to set up and cleaner to visit: The Convo Web Access Link — the public …/twins/your-name address anyone can open without an account — now sits with the other Convo settings in your Digital Twin’s Realtime / Convo section, and shows a scannable QR code beside the link for posters, badges, and screens. The visitor experience is tidier too: guests are no longer asked for their location, the page opens straight into the voice conversation instead of asking whether they want voice or text, and the duplicate close button that sat beside Logout is gone. See Guest Twin Links.Open an embedded twin with an assistant pre-selected (Web SDK): A new assistantId load option lets your embedding page open the Digital Twin in text mode with a chosen assistant already staged — the interface shows it as the pending choice, and the visitor’s own first message is routed to it. Nothing is posted on their behalf: no injected opening turn, no conversation created until they type. Works on text-only twins (voice not required); if the visitor switches assistants themselves, their pick always wins. See the Web SDK guide.Twin Memory: Your Digital Twin can now remember what matters across conversations, and shows you a receipt each time it saves something — click it to see exactly what was kept. Review, correct, export or permanently delete anything it remembers in Settings → Memory. It is additional: your existing User Memory and Shared Memory are unchanged and keep working alongside it. Availability is set per Digital Twin by your administrator. See Twin Memory.Verified learner identity on external MCP tools: When your Digital Twin calls one of your MCP connectors, the request now carries a signed, server-verified identity for the signed-in learner. That means partner tools can safely answer personal questions — a learner’s own progress, quiz-result debriefs, tutoring observations — because the identity comes from the login session and can be independently verified by the partner’s server, rather than from text the AI generated (which a crafted prompt could fake). Admins will find the new MCP Signing Secret on the institution edit screen next to the existing MCP Secret, with copy and regenerate controls: share it with your integration partner through a secure channel, and coordinate before regenerating — rotation takes effect immediately. Existing integrations keep working unchanged; connectors that don’t verify identity simply ignore the extra information.The Agents workspace now fits your phone: The agents dashboard had been squeezing its chat into a sliver on small screens. The sidebar now collapses to a slim rail and expands the same way the conversations panel does — hover to peek on desktop, pin it open, or tap the edge handle on a phone — and it gets out of the way again once you pick an agent. The header reads Agents › agent name as a breadcrumb you can tap to reopen the agent list, replies use the full width of the screen, the message box sits centered beneath the conversation, and chat is now the landing view with Remote Desktop one tap away in the sidebar.Shared twin pages fill the whole screen on iPhone: Public twin links (guest access) no longer leave a colored band at the bottom of the screen on iPhones — the conversation now runs properly edge to edge in Safari and matches the twin’s light or dark theme while the browser’s toolbars come and go.Autonomous Agent settings moved: In the institution editor, the section formerly labeled “Agent” is now Autonomous Agent and sits directly below AI Behavior. The Realtime / Convo options were also regrouped: the kiosk-oriented toggles — Lock full-screen conversation and Start Immersive Mode by Default — each get their own row with their full descriptions visible.Avatar calls no longer fail when an assistant isn’t voice-enabled: Starting an avatar conversation with an assistant active in your chat used to end the call with an error if that assistant wasn’t enabled for Conversation Mode. The call now simply proceeds without the assistant (or with your institution’s default voice assistant, if one is set). To have an assistant participate in voice, an admin enables “Available in Realtime / Convo” on it.A dropped connection no longer cancels your Agent-Mode build: If your network blips, your laptop sleeps, or a proxy times out while an agent build is running, the work now continues on the server and the finished result is saved to the conversation — you’ll see it the next time you open that chat. Long-running builds are bounded by a server-side time limit, and an explicit Stop still cancels immediately. For API users: closing the streaming connection no longer cancels the request; use the documented abort endpoint to cancel explicitly.Cheaper Agent-Mode turns on Kimi lanes: The Kimi models that power coding and planning work in Agent Mode —
kimi-k3 and kimi-k2.7-code — now pass Moonshot’s automatic prompt-caching savings through to you on every turn: cached input tokens are billed at the provider’s actual cache-hit rate (a 10x discount) based on the real cached counts each turn reports, instead of the full input price. Long multi-step build turns, where the model re-reads its accumulated context on every step, get the biggest reduction — the measured build turn cost roughly a third of its former price.Duplicate submissions are now caught: If the same question is accidentally sent again to a conversation while the first submission is still working (a double-click, a reflexive retry after a slow turn), Pria now recognizes it and points you back to the answer already in progress instead of starting a second run — and a second bill.Conversation Mode works from the new chat landing on every twin: The “Tap to start talking” button now starts Conversation Mode correctly on twins without an avatar — you’ll get the microphone permission prompt and go straight into the voice session.Quicker conversation renaming: Hover the conversation name at the top of your dialogue history to reveal a pencil and rename it in place. In the side tray, the conversation menu no longer overlaps the dialogue count.Cleaner pricing view: The Dialogue Report Card’s Pricing tab shows the itemized bill — every lane, model, tokens, and credits — while the explanation of how pricing is calculated now lives in the Plans & Credits guide, including the new Single-Shot and Subagent pricing sections.Keep a conversation in full screen (admins): A new Lock full-screen conversation setting in your Digital Twin’s Voice settings keeps people in the full-screen avatar view for the whole conversation. The collapse button and the Escape key are turned off, and ending the call is the only way out — so a learner in a kiosk or an embedded interview can’t accidentally land back in the chat interface mid-session. Starting, ending and muting the call work exactly as before, and the setting is off by default, so nothing changes unless you turn it on.Gemini 3.7 Flash available: Google’s newest Flash model (gemini-3.7-flash) is now selectable everywhere Gemini 3.6 Flash is — conversation, image and video analysis, summaries, and the meeting/web assistant roles. Built for coding and multi-step agentic work, with the same 1M-token context window and introductory pricing at half the standard rate through the end of 2026.Grok 4.6 now available: xAI’s new flagship model, grok-4.6, can now be selected for Conversation, Image Analysis, and Summary. It keeps the 500K-token context window, image understanding, tool use, and pricing of Grok 4.5, and adds a deeper reasoning tier: setting your Digital Twin’s Reasoning Effort to “max” now unlocks the model’s highest thinking level instead of stopping at “high”. Prompt caching stays automatic, and long-context turns are billed at the standard rate until a prompt actually crosses 200K tokens.Text size controls now resize your conversation: The Smaller/Larger controls in the chat actions menu previously changed some interface elements but left the actual conversation text untouched — which made them feel broken. They now scale your messages, the reply text, and the typing area together, and Pria also respects the text size you’ve set in your browser’s own accessibility settings. Your choice is remembered between visits.Jump-to-latest is easier to find: The double-chevron that scrolls you to the latest message now floats just above the typing area instead of hiding at the bottom edge of the screen, at any input height.Google tools no longer disappear after saving instance settings: Fixed an issue where saving a Digital Twin’s settings — even for something unrelated — could silently switch off Google integrations (Gmail, Drive, Docs, Calendar) for every user of that twin. Affected twins have been restored, and the Google Workspace “Enable Access” toggle now correctly reflects whether the integration is active. If your assistant recently told you a Google tool was “disabled,” try again — reconnecting takes one Google sign-in prompt at most.Branded sign-in links now say when an account has no access: A branded login link (the “Convo Web Access” / invite URLs) paints the Digital Twin’s login page, but signing in with an account that was never given access used to silently drop you on your default assistant with no explanation — which read as “the twin disappeared”. Pria now tells you directly: a dialog names the twin from your link, shows which account you signed in with and where you actually landed, and offers Log Out & Switch Account so you can sign back in with the invited email address. Signing in with an account that has access — or as a platform administrator — now also switches you straight onto the branded twin.Persona mode is now standard: Persona mode has graduated from its experimental phase — the “(Experimental)” tag is gone from the twin settings, and newly created Digital Twins now start with persona mode on, so a new twin speaks as its character from the first conversation: capability and help questions are answered in its own voice, without platform guides or tool lists. Existing twins are untouched — whatever you have set today stays exactly as it is, and the toggle remains available either way.Lock an API key to specific sources (admins): A personal API key can now list exactly which sources are allowed to use it — server addresses (like 52.206.211.0/24 or a single IP) and/or web addresses (like https://app.example.com or *.example.com). Leave the list empty and nothing changes; your existing keys keep working exactly as before. Start with warning mode, where Pria records what would have been blocked so you can confirm the list is complete, then switch enforcement on. Keys are also easier to audit now: the “Recent device usage” list shows the actual tool that connected (for example curl 8.19.0 or node) instead of “Unknown client”.Pria no longer goes blank in privacy-hardened browsers or embedded views: On browsers that block site data or lack certain features — strict privacy settings, some mobile and LMS embedded views — opening Pria could show an empty page: the navigation bar, the admin console, voice mode, shared file and conversation links, and the sign-in redirect were all affected. Every one of these now loads normally; at most a cosmetic preference (like your dark-mode choice persisting between visits) is skipped. Text-to-speech controls likewise no longer break pages on browsers without a built-in speech engine. Behind the scenes, Pria also now reports these rare client-side errors automatically, so problems in the field get found and fixed without waiting for user reports.No more blank screens after an update: If part of the app fails to load — most often right after a new Pria version is deployed while you still have an older tab open — the page no longer goes blank. Pria now retries automatically behind the scenes, and if the content really can’t load it shows a clear “Pria has been updated” message with a Reload button instead of an empty screen. Nothing reloads without you choosing it.Digital Twin video no longer gets stuck on “Loading video…”: Launching a twin in Immersive mode could occasionally sit on the loading spinner forever, with stopping and restarting the twin as the only way out. The video connection now recovers on its own (it retries once automatically), and if it genuinely can’t start you get a clear “The video failed to load” message with a Restart button — plus connection errors are no longer hidden behind the auto-hiding controls in Immersive mode.Class Leaderboard (LXP): Every class now has its own leaderboard, next to the Syllabus and Gradebook — learners ranked by the points and skills they earned in that class only, not across the whole product. Open it from the class card menu on My Classes or from Analytics → Class Reports. Available in every language your deployment is translated into.Grades now reach your LMS gradebook reliably (LXP): When classes are taken through an LMS (Canvas, D2L and others), completed work now reports its grade the moment a student opens the assignment — previously, work finished outside that exact assignment session could silently never reach the LMS gradebook. Administrators also get a Resend grade button to push any learner’s current grade to the LMS on demand, with the LMS’s actual response shown so nothing fails silently.New in Analytics: LTI Grades (LXP): A product-scoped view of every grade exchanged with your LMS — search by learner, filter active vs completed records, hide records with no delivered grade, set how many rows to load, and export the filtered list to Excel. Administrators can resend any grade from the same screen.New in Analytics: Class Reports (LXP): One page listing every class in a product with the numbers that matter — learners, how many completed, average grade, who hasn’t started, inactive members, and last activity — with one-click access to each class’s Gradebook and Class Report.A clearer Analytics menu (LXP): The reporting menu is reorganized by what each page is about — Product, Skills, Classes, Journeys, Learners, and Integrations — instead of the old mixed groupings, and naming is now consistent: what were variously called courses or lessons are Skills everywhere. Menu labels remain customizable per customer, and existing translations are preserved.Sign in with Google lands you in the right place (LXP): Signing in with Google could open your account on the wrong customer/product context — class user lists, lesson lists, and admin pages then looked empty, and your profile picture disappeared. Google sign-in now lands on the same profile a password sign-in would (your most recently used one), with everything attached. If you saw empty lists after a Google sign-in, simply sign in again.Impersonate a learner straight from the user list (LXP): Administrators can now start impersonating someone directly from Manage Users — the row menu gains an Impersonate action, so there is no need to go to the profile screen first. It appears only on users you are permitted to impersonate, and asks for confirmation before switching, because signing in as someone else means logging in again to get back to your own account.Find the right person to impersonate without switching roles first (LXP): The impersonation screen gains a role selector, so administrators who manage several roles can search any of them without leaving the page, selecting a profile, and coming back. Choosing someone in another role now happens in a single step. Super administrators also get an All roles switch for when you do not know which role a user belongs to; type at least three characters and results from every role are listed, each labelled with the customer and product it belongs to.Clearer profile cards (LXP): On the profile chooser, each card now shows Customer / Product so it is obvious which account a profile belongs to, and partner groups are listed alphabetically instead of in an internal order that looked random. Choosing a profile is now done with the Select button only — clicking anywhere on the card no longer switches profile, which removes the ambiguity with the card’s other controls. The delete control moved to the top-right corner as a clearly visible icon, where it no longer overlaps the buttons.Files can now expire automatically when unused: You can set a file longevity — how many days a file is kept without being used in a conversation before it moves to the Trash. Set your personal default on your profile, and (for Digital Twin admins) a separate default per twin under the twin’s settings; both default to 14 days, and 0 keeps files forever. Only files that are actually included in AI search count down, and every time Pria uses a file to answer a question its clock resets — so files you rely on stay put, and forgotten ones clean themselves up. Files show a small “days left” badge as they approach expiry, you can bulk-extend selected files (+7/+30/+90 days or keep forever) or set a specific file to forever from its preview, and expired files land in the Trash (restorable for 30 days) rather than disappearing. Every file already in your vault is unaffected — longevity only applies to files uploaded from now on.Deleted files now go to a Trash Can: Deleting a file — one at a time, in bulk, or by deleting a collection — no longer removes it immediately. It moves to a new Trash view in your Files vault, where it stays for 30 days and can be restored with one click. Restored files come back exactly as they were, including whether they were included in AI search. While in the Trash a file is hidden everywhere: it does not appear in your vault, in collection downloads, in the Knowledge Tree, and Pria will not use it to answer questions. After 30 days it is permanently deleted, and you can empty the Trash yourself at any time to reclaim space sooner.Deactivating a customer account now fully disables its Digital Twins: When a Praxis administrator marks a customer account inactive, every Digital Twin under that account immediately stops accepting new sign-ins and API activity across all channels — web, SDK embeds, LMS/LTI launches, MCP connections, avatar embeds, and content curation. Users who are members of other, still-active twins are guided to switch profiles rather than being signed out. Reactivating the account restores access within seconds, with no per-twin changes needed.AI-curated reading is live again, now powered by Pria (LXP): The recommended-articles carousel on courses and mission briefs stopped refreshing in December 2025 when the third-party service behind it was retired, leaving the same stale articles in place. Curation now runs on Pria’s own search, refreshed daily, and every suggested link is checked to be reachable before it is shown — no more dead links in the carousel. Existing articles were kept, so nothing you had disappears; new ones simply appear alongside them. Search terms are now stored in the LXP itself and can be edited per briefing under Manage Briefings, where previously they lived with the outside vendor and could not be reviewed. Where no terms have been set yet, they are suggested automatically from the course’s own skills and topics. To keep things current without wasted effort, only briefings attached to a course someone has opened in the last week are refreshed; a course that goes quiet keeps its existing articles and resumes refreshing as soon as a learner returns.Email delivery restored (LXP): Fixed a long-standing issue where system emails — password resets, activation links, enrollment notifications, instructor messages, and reminders — were not being delivered. If you requested a password reset in the past and no email arrived, it will work now.Expired lessons can now be un-expired (LXP): If a lesson showed Expired on a course that is not subscription-based, the Extend button silently did nothing — there was no subscription period to extend, and no way to clear the stale date. Extend now removes the expiry on courses that do not expire, so the lesson opens normally again, and the confirmation wording reflects what will actually happen instead of offering to “extend by 0 days”. Existing affected enrollments have been corrected, so anyone who hit this should find it resolved without doing anything.Bulk “Extend subscription” now updates the right enrollments (LXP): In Course admin, extending a subscription for several selected learners at once could write the new expiry date against unrelated enrollments while leaving the selected learners unchanged. It now updates exactly the chosen learners on the chosen course, and reports how many were updated and how many were skipped because they are not enrolled.Tidier archive for busy accounts on My Cohorts (LXP): If your archived cohorts span more than ten partner/program groups, the Archived section now opens as a compact index — every group heading and count visible, tiles tucked away until you expand a group with its checkbox. Searching still finds everything: while a search of three or more characters is active it looks inside collapsed (and unticked) groups too, revealing any match — the group checkboxes grey out to show search is temporarily in charge, and your expand/collapse choices return untouched when the search is cleared.My Progress page loads dramatically faster (LXP): Opening My Progress (competency scores and learning-resource statistics) previously re-computed data for every profile you hold, even though only the selected profile is shown. It now loads just the profile you’re viewing — on multi-profile accounts the page went from about five seconds to well under one.Find any cohort fast on My Cohorts (LXP): The My Cohorts page gains a search bar under the welcome message. Type three or more characters to filter your cohorts live by cohort name, partner, or program — no page reload — with checkboxes to include or exclude active and archived cohorts, a running match count, and a clear “no cohorts match” notice when nothing fits. Reordering tiles by drag-and-drop pauses while a filter is narrowing the list, so a hidden tile can never be shuffled out of place unseen.Archived cohorts grouped by partner and program (LXP): The Archived section of My Cohorts now groups your past cohorts under partner/program headings with a count per group, so it is obvious at a glance where each archived cohort belongs. Each group has its own checkbox to collapse or expand it, and the search filters inside groups too.Duplicate cohort tiles fixed (LXP): If you belonged to the same cohort through more than one enrollment, My Cohorts could show that cohort twice. Each cohort now appears exactly once.Archived cohort hover glitch fixed (LXP): In the Light and Space themes, hovering over an archived cohort made its outline vanish into the background. The outline now stays visible while hovering.Cleaner top bar on small screens: On phone-width windows the top navigation bar now hides the credits counter and the light/dark theme toggle so navigation buttons have room to breathe — especially alongside the Agent button on agent-enabled Digital Twins. Nothing is lost: “Add Credits” stays in the profile menu, joined by a new Theme entry that switches between light and dark mode at any screen size (it flips instantly and the menu stays open so you can see the change). The Agent Desktop button also now appears only while an agent-enabled Digital Twin is selected — it no longer shows on personal accounts.Kimi K3 available (RunPod): Moonshot AI’s flagship Kimi K3 (
kimi-k3) can now be selected as a conversation model, served through RunPod’s OpenAI-compatible endpoint. It brings a 1M-token context window, native tool calling for agentic multi-turn work, and always-on reasoning. Add your RunPod API key platform-wide or per instance to enable it.Integration wizard: Generate ready-to-use embed code for your Digital Twin without writing a line of it. Pick where the twin will run — Canvas LMS theme, a plain web page, a React app, or a Python-served site — configure the widget visually (position on screen, button size, full-screen, auto-open, start-in-Convo, and more, with a live preview), and copy or download the finished file. Visitor mode generates code that works instantly, giving each visitor a reusable guest identity that is cleaned up automatically when dormant; or switch it off to plug in your own sign-in. Find it under Instance Settings → Integration, or via the Integration button on the admin Instances list (which replaces the old per-row SDK copy buttons). The wizard also offers a Claude (LLM instructions) target — a drop-in CLAUDE.md that teaches Claude to drive your twin through the REST API and Swagger — and an Avatar widget target that generates the two-tag talking-avatar embed for any website (theme, launcher or inline mode, and microphone behaviour configurable; requires avatar embed to be enabled on the twin).Voice-first login — Start Immersive Mode by Default: Instance owners can launch visitors straight into the immersive voice conversation after sign-in — including one-click Google/GitHub/Facebook logins, SDK embeds, Canvas, and public twin pages. Turn it on in Instance Settings → Voice → “Start Immersive Mode by Default” (also in the admin instance editor). Visitors see a tap-to-start screen with the twin’s avatar (the tap is what grants microphone access — nothing is heard or billed before it) and can exit to the classic chat at any point; with the switch on, the next sign-in starts in the voice conversation again. Full-screen behaviour follows the existing Avatar Fullscreen setting. Digital Twins with a public guest page had it turned on automatically, so their visitors’ voice-first experience is unchanged.Resume-workspace errors no longer vanish: If resuming a sleeping Agent Desktop workspace failed, the error message could disappear immediately, leaving no indication of what went wrong. The message now stays visible so you can read it and retry.Gemini 3.6 Flash available: Google’s Gemini 3.6 Flash (gemini-3.6-flash) is now selectable everywhere Gemini 3.5 Flash is — conversation, image and video analysis, summaries, and the meeting/web assistant roles. Same intelligence tier with noticeably faster responses and lower output cost.Claude Opus 5 available: Anthropic’s new Claude Opus 5 can now be selected on both the Anthropic Direct API (claude-opus-5) and Amazon Bedrock (global.anthropic.claude-opus-5) for Conversation, Image Analysis, and Summary. Same price as Opus 4.8 with a 1M-token context window, 128K output, and reasoning that adapts automatically to the task. Note for Bedrock: your AWS account must have model access to Claude Opus 5 enabled in the Bedrock console before the Bedrock variant will respond.Mistral’s 2026 lineup added: Four new Mistral models are now selectable. mistral-small-2603 (Small 4) brings reasoning, image understanding, and tool use in one very affordable model with a 256K-token context window, for Conversation, Image Analysis, and Summary. devstral-2512 (Devstral 2) is a dedicated agentic-coding model, also with a 256K window. Audio transcription gains voxtral-mini-2602 (Transcribe 2) with improved accuracy at lower cost, and content moderation gains the updated mistral-moderation-2603.Grok 4.5 now available: xAI’s newest flagship model, grok-4.5, can now be selected for Conversation, Image Analysis, and Summary. It brings a 500K-token context window, image understanding, tool use, and adjustable Reasoning Effort — the first grok-4 family model where your Digital Twin’s effort setting (low/medium/high) is honoured rather than the model always thinking at full depth. Prompt caching is automatic and unusually effective on this model, so long conversations cost less per turn.Super admin role stability: Fixed an issue where a super admin logging into an instance that lists them as its contact email could lose their global super role until manually restored.Clear message when the AI returns nothing: On the rare occasion a model ends its turn without producing any text, Pria now shows a friendly “empty response — please try again” notice instead of an empty reply bubble. Applies across all AI providers.Collection downloads now require access: Downloading a collection as a ZIP now verifies you own the collection or hold file-access permissions on its instance. Previously the download link did not check membership.MCP server credentials no longer shown: Admin screens no longer display a configured MCP server’s stored authorization header. The interface indicates whether a credential is set; leaving the field empty on save keeps the existing credential unchanged.Agent Desktop now one click away for admins: Institution admins can open their Digital Twin’s Agent Desktop straight from a new button in the top navigation bar — previously this screen was reachable only by super-users. The button appears for admins of instances that have Agent Mode enabled and leads to the familiar desktop view with its connect and resume controls. Regular users are unaffected.Mute really mutes in immersive avatar conversations: Pressing the microphone mute button during a full-screen (immersive) avatar conversation now fully silences your microphone. Previously, with certain avatar voice configurations, sounds in the room could still be picked up while muted — interrupting the avatar mid-sentence or even triggering unintended responses. While muted, the avatar now finishes speaking uninterrupted and nothing you say is heard or answered; unmuting restores listening instantly.Documents with emoji no longer break conversations: Questions that drew on vault files rich in emoji or special symbols (for example, reports with emoji section headers or tables) could fail with a provider error — and once a file was affected, every question touching it failed the same way. Pria now handles emoji and special symbols safely, so these files just work. Existing files work immediately; there is no need to re-upload anything.Voice conversations recover cleanly from mid-turn errors: If an error occurred while a voice reply was already speaking, the turn could end abruptly with no explanation. Pria now ends that turn gracefully with a clear message instead of hanging.Consistent model management across every feature: admins now manage the models behind every Pria feature in one place — System → System Models (with per-Digital-Twin overrides where offered). Usage and billing always report the real provider that served each call, and your institution’s own provider keys (bring-your-own-key) are honoured throughout.Claude Sonnet 5 available across Anthropic, Bedrock, and OpenRouter: Claude Sonnet 5 can now be selected on Amazon Bedrock, OpenRouter, and the Anthropic API, with introductory pricing (10 per million tokens through August 31, 2026, then 15). It is now the platform default model. Credit charges track actual provider costs more precisely, and Claude Sonnet 4.6’s maximum response length was doubled to 128K tokens.Turn User Memory on or off — one switch, your call: a new Enable User Memory switch (Your Profile → Knowledge → Memory — plus a quick toggle in the conversation action menu, on by default) gives you full control over personalization. Turned off, Pria stops saving new personal memories and stops using the ones it has — across text and voice conversations — while your existing entries are kept untouched and remain viewable, editable, and deletable in the Memory panel. Your team’s shared instance memory is not affected. See How Memory Handles Your Data for the full data-handling picture.Persona mode for Digital Twins: a new toggle (Instance Settings → Personalization, and the admin institution editor) keeps character twins fully in character. When on, questions like “what can you do?” or “where do I start?” are answered in the twin’s own voice, fully in character, and voice conversations use the twin’s personality throughout. Off by default; existing twins are unchanged until an admin turns it on.AI for Good Summit 2026 concierge — refreshed and sharper answers: Nova, the AI for Good Summit assistant (ai4good.praxislxp.com), has been fully refreshed ahead of the summit (7–10 July, Palexpo Geneva) with the latest official programme — 342 sessions, 1,231 speakers, 141 exhibitors, and 18 topic tracks. Answers also got more complete and better grounded: every speaker profile now includes their LinkedIn and social links and lists their sessions with the date, time, and room; exhibitor pages include contact and social links; and day-by-day and per-track answers are more reliable, so questions like “what’s on Thursday?” or “list all the AI for Health sessions” land on the right, complete answer.Bulk actions across more admin lists: The select-many toolbar now works beyond Users. Instances support bulk Set active / Set inactive / Delete, and Conversation history, Sessions, Assistants, Feedback, AI Models, Entitlements, and SSO providers each support bulk Delete. Tick rows (or Select all on screen), confirm the count, and they’re actioned in one step with a report of how many changed. Checkboxes respect your permissions — rows outside your institution are disabled — and the selection clears when you search or filter.Bulk-manage users — set active/inactive or delete many at once: The admin Users list now has a checkbox per row. Select users (or Select all on screen) and a toolbar appears with Set active / Set inactive / Delete and a count. Confirm before applying, and Pria updates them in one step, reporting how many changed and skipping any you can’t manage. Delete is a soft delete — accounts are deactivated and their email freed for reuse. The selection carries over as you scroll and clears when you search or filter.Two-step verification with an authenticator app — alongside email: Secure sign-in with an authenticator app (Google Authenticator, Authy, 1Password, or any compatible app), in addition to or instead of email codes. In your profile’s Two-step verification settings, choose Authenticator app, scan the QR code (or enter the setup key), enter a 6-digit code to turn it on, and save the 10 single-use backup codes (regenerate a fresh set anytime). Turn on email and the authenticator app together and manage each independently. At sign-in, Pria asks for your preferred method and lets you switch — “Send an email instead” or “Use a backup code” — right on the screen. An administrator can reset your authenticator if you’re locked out.Branded sign-in no longer carries over between users on a shared device: Logging out now fully clears the institution branding tied to your session, so the next person to sign in on the same device lands on the correct portal for their own account. Returning users who arrived through a branded link still see that brand on the sign-in page as before.Set guardrail instructions for your Digital Twin: Instance Settings (and the admin institution editor) now have an opt-in Guardrail instructions field with an on/off toggle. When enabled, your guardrail text applies across every conversation — chat and voice — for all assistants, without editing each one. Load default template fills it with an academic-integrity / Socratic rule (guide students’ thinking rather than complete graded work) that you can edit. It’s off and empty by default, and appears as its own Guardrail segment in the conversation’s context panel.Reasoning effort adapts to each model’s limits: Some advanced models (such as GPT-5.4 Pro) don’t accept the lowest Reasoning Effort settings. If your Digital Twin’s effort is set below the chosen model’s minimum, Pria now raises it for the turn instead of failing the request. Turning reasoning off entirely is still fully respected.Share your assistant with a public web link: Give any Digital Twin a short web address anyone can open — no sign-in needed. In Instance Settings → Invite → Guest Access, set a Convo Web Access Link (availability is checked as you type), then copy and share the generated link (for example
…/twins/your-name). Visitors land on a clean full-screen page that opens straight into a live voice conversation with your assistant, over your Digital Twin’s background, with a light/dark toggle. If voice isn’t available it falls back to text chat, so the page is never a dead end.See how much of your vault is actually being used: The Files Vault Health panel now has a Vault usage chart comparing how many files are in the vault, included in search, actually used (reached by a query), and processing or in error — each bar labelled with its count and share, with a headline percent of included files reached by queries. Each file card also gains a small usage dot — filled when the file has answered questions, hollow when included but never used — so you can spot unused files at a glance.More file types are now indexed: scripts, notebooks, calendars, and contacts: Uploading a shell script (.sh, .bash, .zsh, .csh), a Windows batch script (.bat, .cmd), an R script (.R/.r), a Jupyter notebook (.ipynb), a calendar (.ics/iCalendar), or a vCard contact (.vcf) now works — Pria reads and indexes their content for search instead of rejecting them as “not supported.” For a calendar, event titles, descriptions, and dates become searchable; notebooks are read cell by cell.Turn on the knowledge graph for your personal vault — no reload needed: When you enable the knowledge graph (KAG) for your personal files, the option now appears immediately everywhere it’s offered — including the Rebuild knowledge graph choice in the Reprocess dialog and your Vault Health panel — instead of only after a page refresh.Video scene analysis: Videos now get a visual description of what was shown on screen — not just the spoken words — so “what was shown” becomes searchable alongside the transcript. Admins choose the video-analysis model in System → System Models (a video-capable Gemini model by default) and can override it per Digital Twin under Institution → AI Models.Video and audio files transcribe reliably: Uploads that previously failed with connection resets or a “failed to parse audio file” error — common for .mp4 videos — now transcribe successfully, so their spoken content is captured for search. Silent videos (no audio track) are recognised up front and marked clearly (“File has no audio track.”) instead of failing with an opaque error.Tune search precision and detail per Digital Twin: Admins can set a Chunk Size for each Digital Twin under Institution → RAG and Files — Precise for sharp, fact-based answers, Balanced (the new default) for most content, or Broad to keep more context together — with a companion Chunk Overlap control. New defaults give more accurate answers at lower cost; existing files keep their current setting until reprocessed, and new uploads use the new setting automatically.Knowledge controls in Instance Settings: The Knowledge tab in Instance Settings now exposes two whole-Digital-Twin controls without opening the full admin Configuration page (admin access required). Ignore user personal vault — when on, the assistant answers only from the shared instance library and won’t draw on members’ personal uploads, for more consistent responses. The Chunk Size slider (above) is available here too.Edit a file’s details right in the Asset Info panel: A file’s title, filename, description, authors, dimensions (images), thumbnail, whether it’s included in search, and its visibility (private/public/confidential) are now editable directly in the Asset Info panel of the file preview. Click a field’s pencil to edit inline, flip a switch to include or exclude it from search, or click the thumbnail to change it — every change saves on the spot. You can also enlarge the Asset Info panel with the new double-chevron next to the pin.Edit a file’s text right in the preview: For text-based files — Markdown, HTML, plain text, and code — the preview now has an Edit content button that opens the whole file as editable text. Make changes inline and save, and the file is automatically re-indexed. (Binary files like PDFs and images aren’t editable this way.) A companion Edit Segments button still lets you fine-tune the individual passages the AI uses; open Properties to reach both editors.Editing a file segment now refreshes the knowledge graph too: When you edit or delete a file’s content segment, the change is now reflected in both search and the knowledge graph — previously only search updated.Track your background exports: Admins can now see every “Export all” job under System → Email Exports — its status (queued, running, done, or failed), the number of records in the finished CSV, who ran it and for which Digital Twin, when, and a link to re-download. Failed exports retry in one click, and a column selector lets you choose which columns to show (remembered).Instance Settings shows the correct tool on/off state: The Tools panel in Instance Settings now reliably reflects which tools are enabled or disabled for a Digital Twin, and toggling saves correctly. Previously some disabled tools could appear enabled.Export all — now on every admin table: The full-dataset “Export all” (prepared in the background, emailed as a private download link) is now available across all admin lists — Tools, AI Models, Saved States, Questions, session monitoring, Sessions, Payments, Entitlements, Feedback, Assistants, and Conversation history — in addition to Accounts, Users, and Institutions. Large exports now complete reliably no matter how many records match (a 179,000-row users export that previously failed now succeeds).Tree of Knowledge search finds everything: Searching the Tree of Knowledge now finds any entity by name — not just the handful shown in the default curated view — with a clear loading indicator while results stream in, instead of a premature “no matches.”Tree of Knowledge opens fast for large knowledge bases: Opening the Tree of Knowledge for a Digital Twin with a very large knowledge base — tens of thousands of entities — is now near-instant, where it previously took several seconds. Smaller vaults are unchanged.Reprocess into search and the knowledge graph — your choice: When reprocessing a file you can now tick Rebuild search index and/or Rebuild knowledge graph independently (tick both for a full rebuild). The same choice now applies to bulk reprocessing and to reprocessing an entire collection from its menu, so a whole collection can be rebuilt into the knowledge graph in one action. The knowledge-graph option appears only on Digital Twins that have it enabled.Move a sub-collection out of its parent: Nested collections now have a Move out of collection action that lifts the collection up one level — previously you could drag a collection into another but not back out.Export full datasets from admin tables: Admin list pages now have a CSV button with a dropdown — Export visible downloads the rows on screen instantly, while Export all extracts the entire filtered dataset. Large exports are prepared in the background; you’ll get an email with a private download link when ready (saved to your own vault, not shared). Available on the Accounts, Users, and Institutions lists, with System Models and usage views exporting as standard CSV.Search inside your files — with graph results and inline reading: The Files content search now includes knowledge-graph (KAG) matches wherever KAG is enabled for your personal vault or institution, not only keyword matches. Expand any result in place to read the full matched passage — and for graph-backed hits, see which knowledge-graph entities matched — without leaving the search panel.Users page tabs: The admin Users page now has tabs — Users, Statistics (per-user credit & usage stats, moved here from the History page), and Cap Effectiveness, a new view charting how the per-user 24-hour credit cap performs across instances. The Account, Institution, Account Type, and search filters apply to every tab.Admin list column controls: The Accounts, Institutions, and Users lists now have a Column Selector to show or hide optional fields — credit caps, model and voice configuration, KAG flags, last activity, per-user 24-hour usage, and more — remembered per admin. CSV exports now match the columns you have on screen. The Institutions list adds a per-user credit-cap indicator (Cap Tier, Per-user cap, Per-user 24h cap), and the History list gained optional metric and threat-status columns.Statistics tab — per-user credit-cap monitoring: The admin History → Statistics tab gains four optional columns — Effective Cap, Lifetime Used, % of Cap, and At Cap — surfacing each member’s credit-cap health alongside period usage. A summary line shows mean credits per user (by account and institution) and how many users are at their cap. The new columns are hidden by default; turn them on from the Column Selector.Instance user quotas — admins set per-user limits directly: Admins and managers with the institutions.edit entitlement can now set per-user credit quotas on their Digital Twin — a per-user lifetime cap and a per-user rolling 24-hour cap (the 24-hour cap recovers automatically as recent usage ages out). Set them in the new Quotas tab in Instance Settings or the Credit section of the admin Configuration panel. When the parent Account has caps enabled, the instance inherits them read-only; otherwise it self-configures. Leave a cap blank (or 0) for no limit. See the Credit Allocation & Transfers guide.Usage chart — see how many users were active: Institution admins viewing usage for all users of a Digital Twin (Pricing → Usage) now also see how many distinct users were active in each period, drawn as an amber line over the credits chart with its own axis, matching your chosen bucket size (daily, weekly, or monthly). The running total gains a true distinct-user count for the whole window. Your personal usage view is unchanged.Credit usage caps: Operators can now set per-user credit limits on an account — a lifetime account-wide cap, a per-instance cap (with overrides), and a rolling 24-hour cap. When a user reaches a cap they’re blocked from spending more until it’s raised (or, for the 24-hour cap, until recent usage ages out). Users see a small usage gauge in their gallery and a clear notice when they hit a limit; admins can review usage-versus-cap and how many members are at a limit. Caps are off by default.See your credit usage over time: The Pricing page has a new Usage tab charting the credits you’ve spent per day, week, or month for the selected Digital Twin (or your personal account). Pick a window (7 to 180 days) and bucket size, and the bar chart shows spend per bucket with a running total of credits and messages. Institution admins get a checkbox to switch between your own and all users’ usage for that Digital Twin (hidden for personal accounts). Reading the chart is free.Persona Builder — a guided onboarding interview for every Digital Twin: Personalizing a Digital Twin now runs through one curated 14-question interview — the Persona Builder — grouped into five sections (Basic Identity, Role and Primary Use Case, Audience, Communication Style, Advice Style and Guardrails), identical on the admin onboarding tab and in the create-a-twin wizard. It uses richer question types: selectable tiles for single-choice questions and a personality slider group (a 1-to-7 scale for each of seven traits: tone, energy, detail, feedback, humor, thinking, warmth). Every answer feeds the twin’s Instructions via Generate Prompt. Required questions must be answered to continue; the email question prefills from your account (and never changes your login), and the twin-name question stays in sync with the instance name. See the Onboarding Questions guide.Onboarding popup now skips already-personalized twins: The first-time prompt to run the Persona Builder no longer appears for Digital Twins that already have Instructions. It still shows for un-personalized twins and continues to yield to the Canvas sign-in prompt where that applies.Read responses aloud with a natural voice: Pressing Speak on a response can now play it with a natural-sounding voice from your Digital Twin’s text-to-speech model instead of the basic browser voice. Institution admins turn it on per Digital Twin in Instance Settings → Voice → Natural Voice and pick the read-aloud voice there; personal accounts get a Use Natural Voice toggle in Your Profile → Conversation Preferences with a Playback voice picker. The chat’s [+] → Audio menu shows the current state and jumps to the right settings. Every voice picker has a Preview button that plays a short sample (previews are free). Generating audio consumes credits based on response length; replaying the same response is free.Claude Fable 5 — Anthropic’s newest flagship model: Claude Fable 5 is now available as a conversation, image-analysis, and summarization model, on both the Anthropic API and Amazon Bedrock. It brings always-on adaptive reasoning, a 1M-token context window, and 128K-token responses. Select it in Admin → AI Models (claude-fable-5 on Anthropic, global.anthropic.claude-fable-5 on Bedrock — the Bedrock variant requires an AWS account data-sharing opt-in; see the AI Models guide).Zero Data Retention — clear compliance signals in the model selector: Praxis AI ships only models that comply with Zero Data Retention (ZDR) — your prompts and responses aren’t stored by the model provider — and the model selector now says so. The one exception is clearly flagged: Anthropic’s Mythos-class Claude Fable 5 carries a red warning and a “30-day retention” marker, because Anthropic retains prompts and outputs for these models for 30 days (effective June 9, 2026). The warning links to Anthropic’s retention notice, and the AI Models guide has a new ZDR section.Lower AI costs — smarter prompt caching across every provider: Repeated conversation turns now cost less across all six AI providers (Anthropic, Bedrock, OpenAI, Google, Mistral, and xAI), with the savings passed on as larger credit discounts. The dialogue info card shows cache activity per response, and a new copy button lets you copy all its metrics as JSON. In Conversation (voice) mode you can also tap the new full-screen (TV) button to put the assistant’s avatar front and center, filling the entire screen — great for presentations, kiosks, and hands-free use. It works from the compact widget and the expanded view, goes true full-screen on desktop and Android (fills the screen on iPhone/iPad), and you can exit any time with the on-screen button or the Esc key.Praxis Shield — threat monitoring for your Digital Twins: Institution administrators get a new Threats panel. Conversations are scanned automatically in the background, and potentially abusive or concerning activity surfaces as incidents with a severity, status, and categories, plus a detail report of the relevant conversation context. A built-in triage workflow lets reviewers add notes, progress status, and see the full action history; confirmed abuse can suspend the user’s membership in that Digital Twin (reversible any time), and harmless flags can be marked false positive. Filter incidents by account, Digital Twin, status, or minimum severity. Scanning is invisible to end users and never blocks or alters their experience. See the new Praxis Shield admin guide.Praxis Shield — shareable incident reports: Every incident has a read-only report you can share by link with another administrator of the same Digital Twin. Reports are login-gated and only open for admins who manage that institution, so sensitive context stays in the right circle.Admin console — cleaner navigation on large screens: The admin menu can now collapse into a slim icon-only rail (your preference is remembered), and multi-tab dashboards share a unified layout where the tab bar becomes a collapsible left sidebar and only the content panel scrolls — so headers and navigation stay in view while you work through long lists.Embed your Twin’s talking avatar on any website: A new Avatar Embed Widget puts your Digital Twin’s voice-enabled avatar on an external site with a single script tag (or a React component). Choose a floating launcher or an inline embed, light or dark theme — visitors talk to your Twin without leaving your page, and only a public embed key is required. See the Avatar Embed Widget guide.Admin — run background AI jobs on cost-effective Bedrock models: Background AI jobs (knowledge-graph extraction, threat detection, and summarization) can now use low-cost open-source GPT models on Amazon Bedrock (gpt-oss-120b, gpt-oss-20b, and the safety-tuned gpt-oss-safeguard pair). Several inexpensive models (Amazon Nova micro/lite/pro, Gemini Flash-Lite, and others) are now also selectable for summarization, giving you a price/quality ladder. Each model shows its price per million tokens and capability badges (structured-output, code, vision, reasoning) so the trade-offs are clear at a glance.Admin — see credits consumed per member, per instance: Each membership now tracks the lifetime total of pooled credits that member has drawn from the instance, shown as a Credit Consumed column in Admin → Users → Membership and on the Entitlements list (and in the Entitlements CSV export). It complements credits awarded — handy for usage reporting and spotting unusual consumption. It applies to instances that pool credits across members; personal accounts are unaffected. On the Membership tab it replaces the rarely-used Award Date / Award Amount columns.Admin — merge two user accounts into one: Admins can consolidate a duplicate or vanity-email account into another (for example folding an old personal-email login onto a member’s work account). From the Users list, the new Merge into… action lets you pick the surviving account, see a preview of exactly what will move (and any overlaps or billing caveats), then type the source email to confirm. On confirm, Pria moves all of the source’s content, history, and memberships to the target, sums their credits, removes the source’s logins, and permanently deletes the source — recorded in an audit log. Requires both edit and delete permission on the users involved.Admin Conversation History — show or hide user-deleted conversations: The admin History list now hides conversations a learner removed from their own history by default, with a Show user-deleted toggle to bring them back when needed. Revealed records carry a small deleted badge. A learner’s delete is still respected everywhere it was before — this only changes what an admin can optionally surface.Conversation exports & history — assistant replies no longer appear blank: Downloaded conversation data (including the GDPR-style data export) and the admin history detail view now show the full assistant reply text, fixing cases where streamed replies previously appeared empty.Privacy — stronger isolation of personal files and memory: One member can never read, list, or modify another member’s personal-vault or confidential files, and can’t overwrite or clear another member’s or the organization’s shared memory — those stay owner- and admin-managed. Genuinely shared (instance) files still work across an organization as before. The unused “list all members’ files” option was removed.New AI model — Grok Build 0.1 for coding: xAI’s Grok Build 0.1, purpose-built for agentic coding (web development, debugging, and multi-step build tasks), is now selectable as a conversation model. It’s fast and cost-effective, reads images such as diagrams and screenshots alongside your code, and always reasons through a problem before answering. Admins pick it as a Digital Twin’s conversation model in Instance Settings → AI Models.Files — mark a whole collection confidential in one click: A collection’s ⋯ menu now has Make All Confidential and Remove Confidential alongside Make All Public / Make All Private. Making a collection confidential lets Pria keep using those files to answer but stops it from citing them — applied to every file in the collection and its sub-collections at once. (Confidential and public are mutually exclusive, so this also removes public access.) Remove Confidential reverses it so those files can be cited again.Personal knowledge graph — now available to everyone (opt-in): Building a knowledge graph from your personal files is no longer limited to super users. A new Build a knowledge graph from my files toggle in Your Profile → Conversation lets any user opt in, so search can connect related ideas across your documents rather than just match keywords, and your chat Knowledge → Sources menu gains the RAG + KAG Fusion option (shown when a knowledge graph is available on your personal vault or current Digital Twin). Files shared through an organization keep using that organization’s own setting. Turning the toggle off keeps any graph already built, so re-enabling is free.Personal knowledge graph — process the files you already have: New uploads are added to your knowledge graph automatically once you opt in. For files uploaded before enabling it, a panel under the toggle shows how many aren’t in your knowledge graph yet and an Add my existing files button to process them, with a live progress bar and a Stop control. Files you add next always take priority, so turning this on never slows new uploads.File Preview — knowledge-graph usage itemized per file: Each file now tracks its knowledge-graph processing cost separately from standard search-indexing, so the File Preview shows what each step cost for that file.Admin — KAG Fusion is now a self-service setting (no longer Praxis-managed): Knowledge-graph fusion can be configured directly in the admin console instead of requesting it from the Praxis team. Accounts gain a Knowledge Graph (KAG / RRF Fusion) section with two cascading controls — Enable KAG for all instances and Allow instance-level KAG toggles (the first takes precedence). Instance Settings → Ingestion has an Enable KAG Fusion for your Digital Twin toggle that any admin with instance edit rights can use, and it shows when the setting is inherited from the account. Admins can also set a specific user’s personal-vault KAG from Admin → Users → Knowledge. Each surface links to a More info on KAG reference and notes that building the graph improves answers on complex questions but uses more credits than plain embedding.Admin — process a Digital Twin’s existing files into its knowledge graph: Like the personal-vault control, Instance Settings → Ingestion now shows — once KAG Fusion is on for the Twin — how many of that Twin’s files aren’t in its knowledge graph yet, with an Add existing files button, a live progress bar, and a Stop control. Any admin with edit rights on the Digital Twin can run it.Files — knowledge-graph coverage in Vault Health: The Files modal’s Vault Health panel now has a Knowledge Graph (KAG) section that tells you whether all eligible files are in the knowledge graph and, when some aren’t, gives an Add existing files button with a live progress bar and Stop to backfill them right there. On your personal vault, if KAG isn’t on yet, it offers an Enable for my personal vault button that switches it on and starts processing in one click. It appears for the Personal and Digital-Twin vaults, and is hidden in the combined All-Vaults view.Tree of Knowledge — no longer super-only: The Files → Tree display toggle — a visual map of your knowledge graph (pillars, entities, relationships, and the source files behind them) — is no longer limited to super users. Anyone with a knowledge graph can switch between Files and Tree: users who belong to a Digital Twin, or who’ve enabled KAG on their personal vault. The vault picker inside the Tree only offers vaults you can reach, and the tree only shows knowledge built from files you can already access.Admin — Accounts page polish: The General tab is grouped into clear sections (Account, Managers & Domains, Knowledge Graph, Privacy Policy); the Privacy Policy Consent Form is now a Markdown editor with a live Preview; the number of linked instances shows on the Instances tab (e.g. Instances (18)); and the unused Qualification field was removed.Image generation — request larger or differently-shaped images: When you ask Pria to create an image, you can now request a specific size or shape — a wide landscape, a tall portrait, an aspect ratio like 16:9 or 9:16, or a higher resolution where the selected image model supports it. The default is still square, and if you ask for a size a model can’t produce, Pria uses the closest available size instead of failing.Image generation — faithfully enlarge an existing image: You can now ask Pria to upscale an image you already have (for example “enlarge this 2x”), increasing the resolution while keeping the picture exactly as it is — same composition, colors, and detail. This works with any configured image model. (It’s a clean enlargement, not AI “enhancement,” so it won’t invent new detail.)Backgrounds — set them from a picture or your webcam: The light- and dark-mode background images can now be set like a profile picture — click the background thumbnail to upload a file or capture from your webcam, crop it to a wide 16:9 frame, and save. It’s available in Instance Settings → Personalization → Appearance and, when editing an organization, on every picture in the Personalization tab. Webcam capture requests HD so backgrounds stay crisp, and you can still paste an image URL directly.Conversation History — see what was retrieved during each turn: Every history record now shows a RAG Search Results panel — alongside Thinking Details and Tool Details — listing the file segments Pria pulled in for each turn (file name, relevance %, and a preview of the text). It appears in the chat list, the public history viewer, and the admin History inspector. Confidential segments owned by another user in your institution show a 🔒 with the preview redacted. History pages stay fast — the full detail loads only when you expand a section.Conversation History — unified detail panels with per-entry copy: The Thinking, RAG, and Tool detail panels now share one consistent look everywhere they appear (chat list, single-record viewers, admin inspector), and each entry has a copy-to-clipboard button so you can grab a reasoning round, a retrieved chunk, or a tool response in one click. Tool Details now loads full arguments and responses the moment you open it — no more follow-up click. Live streaming Thinking matches the saved view.File Preview — Processing timeline for RAG and KAG: The Asset Info → Processing section of the File Preview modal now splits into two color-banded blocks — 🔵 RAG · embeddings and 🟣 KAG · knowledge graph — each with a step-by-step timeline (Enqueued → Started → Completed) showing durations and timestamps, a ✓ on finished steps, a pulse on the step in progress, and a red ✗ if a run failed. Each model row shows a provider chip (e.g.
text-embedding-3-small [openai]) so admins can see which model handled the file. A combined Total tokens and Cumulative time pair appears when both phases have run, with per-phase figures in parentheses. The Properties Uploader row was renamed Owner to match the rest of the platform.File Preview — live KAG progress: While a knowledge-graph job is running, the Processing > KAG block now shows a Progress: N% row and the current step inside the timeline. The same live step appears for RAG while a file is being processed; once finished, the completed step is the only signal needed.Hide live Thinking / Tool indicators without losing the audit trail: Two new toggles in Your Profile → Conversation (and on the admin instance edit page) let admins hide the live streaming surfaces on their own. Display Thinking Execution controls the 💡 reasoning block shown in real time; Display Tool Execution controls the inline ⚙️ Running… / 🔧 Ran ✓ indicators. Turning either off keeps the live conversation clean while the separate Display Thinking Details / Display Tools Details toggles still control what’s saved to History. Both default to on.File Preview — redesigned segment editor: Entering Edit mode now hides the Asset Info sidebar so the segments editor uses the full modal width, with Title, Summary, Rights, and segments in one clean top-to-bottom form under File Properties / RAG Segments headers. A new Rights group lets admins set a file Private / Public and, for instance or account vaults, flip a Confidential switch (locks it for other members) — each saves immediately and the file-grid badges stay in sync. + Add Segment is now a full-width button after the last segment, and each segment has its own Update action (saving just that one) next to Sanitize / Delete. Typing into a segment now reliably persists. On narrow screens the segment header collapses to icons and a compact character counter so it stays on one row. The unused Tags / AI Tags rows were removed.Voice dictation — mic stays available after transcription: The mic button no longer disappears once your dictated text lands in the input. Click it again to add more — each new dictation is appended (with a separating space) so you can build a longer message in several passes. After a transcript comes back, the caret jumps to the end so you can keep typing or hit Enter right away.Files — Properties and Edit Segments merged into File Preview: The old separate Properties and Edit Segments dialogs are now one File Preview window, so you no longer switch screens to view metadata and edit search segments. A collapsible Asset Info sidebar groups Properties / Processing / Knowledge — RAG segments, pinning open or expanding on hover like the Pria and Files-search sidebars. The Edit Segments action opens the preview directly in edit mode; Title and Description gain inline pencil edits. Copy file name and Copy link buttons sit next to Download, and fullscreen now expands the whole window. Action dialogs (Reprocess, Swap Asset, Delete) reliably layer on top.Files — Swap Asset (replace a file without rebuilding its index): A new Swap Asset action in the file’s ⋮ menu replaces the underlying file while keeping its existing knowledge (embeddings, summary, title) and its published URL — so anything already linking to the file keeps working. Useful when the source file was lost but the indexed content is still good, or to refresh the downloadable copy. Supply a new file or a URL for Pria to fetch. If you want the new content searchable, run Reprocess Content afterward. The action is hidden when a file is missing from disk with no embeddings (Delete is the only option there).Files — Vault Health pills are now the filter: The Vault Health pills at the top of the panel (needs reprocess, errors, needs delete, missing, unoptimized, stale >90d, never used) now act as filters — click one to scope the listing to that category, click again to clear. The Vault Summary shows a single Health: <grade> chip (A–F) that jumps to the Vault Health section. A new needs delete category flags files missing on disk with no embeddings, where Delete is the only valid action.Files — sidebar polish, lightbox maximize, RAG tuning, Tree of Knowledge: A second pass on the Files modal. The sidebar is now a pinnable rail (pin to keep open, or hover to expand). Search became a Search Type control (File Name / AI Content) with a multi-line box and a dedicated Search button — typing no longer auto-fires, and a clear-X resets the query. A new All Vaults checkbox searches your full scope; both preferences persist. Content-search results show a “found in” chip with the vault and collection; clicking a hit jumps straight there. Admins with edit rights get inline RAG Tuning sliders (Search Affinity %, Max Articles) that re-run the search on save. Tree of Knowledge is promoted to a top-level Display selector (Files vs Tree). The Lightbox gained a Maximize / Restore toggle and goes full-bleed on narrow screens. File status pills were replaced by a small icon strip (Included / Public / Confidential), and resizing the file list is noticeably smoother.Files — content search across your IP Vault: The Search box gained a name/content toggle. In content mode, press Enter to search across the same files your assistants use (your personal vault, current institution, and account-shared files) and get a scored-snippet results panel. Each card shows the file name, a relevance badge, and a highlighted 200-character snippet; confidential files owned by another member show a row and score but no snippet (🔒 Private content matched — open the file to view). Clicking a result selects the file and snaps to it. A ”← Back to N search results” banner lets you flip back without re-running. Limited to 20 searches per minute.Files — responsive sidebar on large screens: On screens ≥1024px, the Search / Display / Sort controls now stack in a 240px left rail beside the file grid instead of a toolbar above it, with Add Files pinned to the bottom. Smaller screens keep the horizontal toolbar. The Display, Sort, and search-mode controls all adopted the same look as the View Mode switcher, with proper screen-reader labelling.Knowledge picker — clearer Retrieval Mode + Search Only: The Knowledge submenu (in the [+] menu and the Knowledge tab of Your Profile) is reorganised into a three-option Retrieval Mode — Disabled, RAG Only, RAG + KAG Fusion — plus a separate Search Only checkbox you can combine with either mode to return raw vault results without an AI rewrite. The old “KAG Only” label was misleading (KAG always augments RAG), so the new copy makes the relationship clear. The Knowledge row badge reflects the active state — Disabled (red), Search Only (amber), RAG + KAG (blue). Existing profiles carry over with no migration. The glossary gained KAG and Knowledge Modes entries.Voice dictation — in-place recorder with live waveform: Tapping the mic in the chat input now opens a dictation widget right where the textarea was — a pulsing red dot, an MM:SS timer, a live waveform, and Cancel / Confirm buttons. Recording stops automatically at 2 minutes (with a 10-second warning). It uses your institution’s configured speech-to-text engine (Voxtral / Whisper / etc.) rather than the browser’s, so accuracy and language coverage match the rest of the platform. Nothing is stored after transcription — no upload, no history, no embeddings. Works in Chrome / Firefox / Edge / Safari. On Confirm, the text is appended to whatever you’ve already typed.Your Data Rights (GDPR controls): A new Your Data Rights section at the bottom of My Profile → Your Data lets you see, export, and delete the personal data Pria holds about you. A stats card shows your personal dialogues, vault files (with total size), and memory entries. Export my data builds a ZIP (README, dialogues.csv, uploads.csv, memory.csv, plus your original upload files) and emails you a signed download link valid for 30 days. Delete my data opens a confirmation where you choose to forget your Dialogues, your Personal files in Vault, or both. Everything is scoped to your personal data — institution and account data is governed by the institution’s own retention policy and untouched. Exports are limited to one per hour, and emailed ZIPs are purged after 30 days.Files panel — full redesign: A top-to-bottom polish of the IP Vault Files modal — Google-Drive-style cards, a horizontal collections strip above the grid (drag a file onto a folder to move it), a redesigned File Preview with an Asset Info sidebar and lazy-loaded RAG content, and a Grid / List / Table view switcher. The Lightbox scales to the full viewport on wide screens. A new Vault Health header shows an A–F grade plus an attention dot for the most pressing follow-up, and a new “Needs reprocess” category flags files that are Included but have no embeddings (common after an embedding-model switch). The bulk action bar is now sticky. Drag files onto a vault tab (My Files / Instance / Account) to move them between vaults. The new table view (Files and Assistants) supports client-side sort, infinite scroll, bulk select, resizable columns, and a Columns menu (Source, RAG Hits, Last Used, Owner, etc.) that persists per user. The Properties dialog was rearranged into clean per-section cards.Collections — custom colors: When creating or editing a collection (or sub-collection) you can now pick an accent color from ten swatches or any custom hex. The folder icon renders as a solid tinted shape across the strip, list, and card views. Pick the ∅ option to revert to the default outline icon.gpt-realtime-2 is now the default voice model: OpenAI’s newest realtime voice model replaces gpt-realtime-1.5 as Pria’s default for Convo / Speech-to-Speech mode. It brings GPT-5-class reasoning with configurable reasoning effort, a 128k context window (4× the previous default), stronger instruction following, and translation across 70+ languages. Reasoning tokens per turn are now tracked separately as Thinking Tokens in the usage card. Earlier models (gpt-realtime-1.5, gpt-realtime, gpt-realtime-mini) remain in the picker. Note: gpt-realtime-2 audio pricing is materially higher than -1.5 — review your Pricing tab if you run at scale. Institutions that had explicitly chosen -1.5 stay on it; only those without a chosen model move to the new default.OpenAI Realtime endpoint migration (GA): OpenAI realtime voice sessions now use OpenAI’s GA endpoints instead of the deprecated beta ones. The change is transparent to end users — same voice, VAD, and tools — but network admins should ensure outbound HTTPS to api.openai.com/v1/realtime/client_secrets and api.openai.com/v1/realtime/calls is allowed; the legacy /v1/realtime/sessions path is no longer used.Voice agent now sees account-shared files: When you ask the voice agent “how many files do I have?”, it now includes files shared across your organization (the Account vault), not just your personal and institution files. These files are labeled Owner: Account so you can tell them apart in tool-call traces. Previously they were silently missed even though the file modal showed them.Live transcripts during voice sessions: The character-by-character transcript under the voice indicator now works for OpenAI realtime sessions (previously it showed only the final transcript at the end of each turn). Your own speech also appears live as You: … as soon as it’s transcribed.Thinking Tokens — visible in every history surface: Reasoning-model output is now reported separately from the visible answer. Everywhere token counts appear — the inline usage card, the per-message info card, My Profile → History, and the Admin → Histories detail modal and table — a new Thinking value shows when the conversation used a reasoning model. OpenAI (Chat Completions, Responses, and Realtime GA) and Google Gemini report this; Anthropic / Bedrock / Mistral / xAI don’t yet expose a separate count, so it stays 0 for those. Thinking text blocks are still saved as before.Google Cloud — revoke before changing services: Both the personal Google Cloud panel (My Profile → Google Cloud) and the institution Google Workspace integration (Admin → Instance → Configuration → Google) now lock the service toggles while a token is active. To add or remove Gmail / Drive / Calendar, click Revoke Access (or Disconnect) first, then re-authorize with the new selection. An expired personal token shows a Re-authorize Services button. A short hint explains the flow so the disabled state isn’t ambiguous.Welcome / Canvas-auth popups dismiss reliably: The “Welcome to your new AI Digital Twin!” prompt and the Canvas-auth recommendation popup now close as expected when you click Later, Yes, Enable, or Dismiss, instead of occasionally lingering after saving.User Guide & Admin Guide open in a new tab: Clicking User Guide or Admin Guide (from the [+] menu under More, or the sidebar buttons) now opens the documentation site at docs.praxis-ai.com in a new browser tab instead of an embedded panel that rendered empty.Logout — branded login page preserved: When you reach Pria through a branded URL (e.g. /login?brandingId=<id>) and log out, the post-logout login page now keeps your brand assets instead of falling back to the default Pria look.Logout — cross-session state leak fixed: Logging out and signing in as a different user no longer surfaces the previous user’s Digital Twins in the top-navigation switcher. Entitlements, user-institutions, and in-flight requests are now cleared on every logout, and stale requests fired during logout no longer surface as 403 console errors.Instance Settings — Voice tab, explainers, and polish: The instance-settings lightbox gained a dedicated Voice tab (between Personalization and Conversation) with the Convo / Speech-to-Speech model picker, per-provider voice selectors (Gemini Live, xAI, OpenAI Realtime — voice, VAD eagerness, noise reduction, transcription language), and a shortcut to the full admin models page. Retrieval is now called Knowledge and opens with a collapsible explainer covering IP Vault security, RAG, and KAG Fusion; matching About tools and About assistants explainers were added. The Conversation > Behavior card gained Enable Moderation and Cite Sources toggles; Tools Guidance and a new Max Tool Result Size (default raised to 60,000 chars) moved to a Tools tab card. Personalization gained an Enable Location toggle and richer descriptions. The Billing tab hides Add Credits when Disable Add Credits is set, the Invite tab gained a Manage Members & Entitlements shortcut, and the Identity tab was renamed General.Personal Settings — descriptions + Knowledge Sources cues: Each Personalization and Security toggle now carries a short description (Dark Mode, Use Location, Enable STT, Knowledge Sources, Must Change Password, Reset password on SSO Login). The Personalization tab shows the current retrieval state at a glance — an amber dot for Search Only, a red dot when Knowledge Sources are disabled. The Google Cloud entry shows a green “Connected” dot when your personal Google token is valid or the instance has Google OAuth enabled.Action menu polish: Item labels are now consistent Title Case — Audio Notes is New Audio Note, Create new assistant… is New Assistant…, List Assistants is Manage. The Knowledge Sources submenu’s active dot uses the same tones as the Personalization tab (amber for Search Only, red for Disabled). The Switch assistant for next message picker is now a clean filtered list that keeps the current assistant with a “Current” pill so you can clear an override in one click. The next-assistant chip sits to the left of the send button.Instance Settings — Invite & Billing tabs, plus Conversation toggles: The instance-settings lightbox gained two tabs. Invite generates a join QR code and copy-to-clipboard invite link for new members, plus a branded login link for existing ones. Billing shows the instance credit balance with an Add Credits button (to the Pricing page’s credit tab), a note on pooled vs. non-pooled credits, and a payment-history table for this twin. The Conversation → Behavior card added Enable Moderation and Cite Sources toggles, and Tool / Assistant rows now show their description under the name.Profile & Instance Settings refresh: Both surfaces were reorganized. In My Profile, Profile is now General, Preferences is Personalization (with the Dark Mode toggle and a new Appearance card), and a new Billing tab shows your personal credit balance, an Add Credits button (to the Pricing credits tab), a payment-history table, and a note explaining when personal credits are used. In Instance Settings, the Features tab is now Personalization and carries a new UI Theme selector (User Directed / Dark / Light), the Display Thinking Details / Display Tools Details toggles, and the Guest UI toggle alongside the file-upload / audio-notes / clipboard toggles.Dialog Usage Details — in-place card: Each message header now shows a bar-chart icon on hover next to the timestamp. Clicking it swaps the message body for a color-coded usage card with the model name, a provider chip, full timestamp, token counts (total / input / completion / baseline), cache discount, credits, retrieval timing, tool names and count, thinking-block count, latency, and request duration. Click again to return to the message. This card is now visible to every user (previously admin-only), and all row colors meet WCAG AA contrast in light and dark mode.Memory Panel — filter by instance: When your memory spans more than one digital twin, My Profile → Memory now shows a searchable instance picker beside the namespace and scope filters. Each option shows the institution’s logo, ainame, and display name, with a “Personal” bucket for entries not tied to an instance. It defaults to your current instance and stays hidden when your memory lives in a single instance.Push-to-toggle microphone: The dictation mic now stays on until you click it again. Previously it stopped after a short pause and needed a fresh click for the next phrase. Now one click keeps it running through pauses and silences, and phrases accumulate in the input as you go — it stops only when you click it off or submit.Admin Histories — column selector + User Statistics tab: Two upgrades to Admin → Histories. (1) The Records and Usage tables each have a Columns dropdown — toggle any column off to declutter, saved per-admin and dropped from the CSV export too. (2) A new User Statistics tab shows one row per user per day / week / month with query count and credits used, respecting your toolbar filters. Toggle the period and click CSV for a per-user usage extract. Results cap at 5,000 rows with an amber banner when truncated.Gemini 3.1 Flash Lite is now GA: Google promoted gemini-3.1-flash-lite out of preview, and the model picker on Personalization → AI Models now offers the GA identifier. Institutions on the preview model are upgraded automatically on the next boot — same architecture and parameters, no behavior change. The preview identifier stops accepting requests upstream on May 25, 2026; pick the GA entry manually if you ever switch away and back.Personal API Key usage: Admins can now see which devices have used their personal API key. Open My Profile → Developer — API Key and expand Recent device usage for per-device rows with browser/OS family, IP subnet, optional client name/hostname (set via x-pria-client-name / x-pria-client-hostname headers), first/last seen, and a call count. Rotating or revoking your key wipes the list. Forget device removes a row but doesn’t invalidate access — regenerate your key to do that. Records auto-expire 180 days after a device’s last sign-in.Cross-twin-safe conversation operations: Renaming, deleting, and AI-summarizing a conversation now scope strictly to the active Digital Twin, so an action in one twin can no longer affect a same-numbered conversation in another twin you belong to — most noticeable for mobile users who switch twins often. Web is unchanged. Mobile and other API clients can pass institution explicitly to /api/user/clearHistory, /api/user/renameHistory, and /api/user/summarizeHistory (an ObjectId to target a twin, null for personal history, or omit for the previous fallback). Membership is verified server-side (returns 403 if you’ve left the twin). The course list also restores the most recently used assistant per conversation.Email-based two-step verification (admin opt-in): Admins now see a one-time prompt after login to enable email-based two-factor authentication. When on, password and OAuth logins from a new device or location trigger a 6-digit code emailed to the account, valid for 5 minutes. Trusted devices stay trusted for 7 days on the same network — paste-from-clipboard and auto-submit on the 6th digit, no button to click. Manage devices or toggle MFA from My Profile → Two-step verification; admins can override per user via Users → Edit → Credentials. Standard users can opt in but aren’t prompted yet. After 5 wrong attempts a code invalidates; after 5 burned challenges in an hour the account locks for 10 minutes. iOS Safari + Mail surface the code in the keyboard autofill bar.Reliable long-audio transcription: Long audio uploads (lectures, interviews, recorded meetings) no longer fail with “audio duration exceeds maximum” when compressed files were small in size but long in playback. Chunks are now sized against both limits, so a heavily compressed 75-minute recording splits into pieces the transcription model accepts. Files that previously failed can simply be re-uploaded.Reliable large text uploads: Large plain-text uploads (HR / medical / administrative exports, CSV/TSV dumps over ~16 MB) that previously failed during processing now succeed — the per-upload extracted-text ceiling was raised to ~60 MB. Files that previously failed can be re-uploaded. Alongside this: ffmpeg media-split errors now show the actual ffmpeg message, and large .tif / .odt / .ods / .odp uploads are no longer rejected.Reliable PDF export from your vault: Pria can now convert an existing markdown file in your vault to PDF without re-reading it — just point Pria at the source file by URL and the conversion happens in a single step. Together with recent fixes for accent and emoji rendering, French and multilingual letters now export correctly with proper bold headings and rendered emoji. Saving a file to a .pdf path renders the Markdown content directly.Simpler save: When you ask Pria to save a file, the filename and extension are now a single path (e.g. report.md or letter.pdf). This removes a class of bug where the requested format and the extension could disagree.Larger content reads: The default cap when summarizing scraped pages, transcribed media, and tool results is now 60,000 characters (up from 40,000) — useful for longer letters, articles, and transcripts. It remains overridable per institution via the CONTENT_MAX_SIZE environment variable.Memory Parameters page: Pria now exposes the running memory it uses to remember context about you — namespaces, key/value pairs, and per-instance shared knowledge — via a new My Profile → Memory entry (also under Pria → Settings → Personal → Memory Parameters). Personal entries are private to you; shared entries are scoped to a digital twin and visible to that instance’s members. You can edit values and descriptions, toggle the shared flag, delete a row, or bulk-delete a filtered selection with a preview before confirming. Members with the institutions.edit entitlement can also edit and clear shared rows for the instances they administer; everyone else sees those rows read-only.Admin Memory Inspection: Admins with the users.edit entitlement can inspect and manage any user’s memory from Admin → Users → Edit user → Memory. Memory is grouped by instance, and admins only see entries for instances they hold users.edit on. The same edit / delete / bulk-clear controls apply, with an authorization re-check on every change.Instance Memory Inspection: A new Instance Memory tab in Admin → Instances → Edit instance shows the shared memory scoped to that instance, with the author email linking to that user’s Memory tab when the admin holds users.edit on their institution.Act As (impersonation): Admins with users.edit can now impersonate a user from the Edit User modal via a new Act As… button. A consent dialog requires a written reason (up to 500 characters) and confirms the admin will see the user’s personal data and that every action is recorded as theirs. While impersonating, an amber Acting as <email> pill stays pinned in the top navigation, the rest of the admin console is unavailable, and Stop restores the admin’s session and returns to the Edit User modal. Sessions auto-expire after 1 hour, and every start and stop is recorded in an audit log with the reason, IP, and user agent. Self-impersonation, Praxis AI staff targets, and nested impersonation are blocked.Display Thinking Details (admin toggle): Pria now collects model reasoning from the providers that expose it (Anthropic, AWS Bedrock, Google Gemini, OpenAI summaries, xAI) and shows it in a collapsible 💡 lightbulb block above the response. Mistral doesn’t expose thinking, so no lightbulb appears. A new Display Thinking Details checkbox on the institution Personalization page (next to Display Tools Details) controls whether thinking is streamed, shown, and saved to History — defaults to on. Personal users always see thinking; institution accounts inherit the toggle. Streaming clients receive a new
THINKING event — see the Streaming docs for the payload shape. History pages load the full thinking body only when you open the lightbulb, keeping list responses lean.Safer bulk delete in IP Vault: The Select All button in the Files panel now respects your active filter — searching scr_* and clicking Select All picks only the matching files, not the whole vault. The bulk-delete confirmation lists every collection and file to be removed in a scrollable preview so you can review before confirming. Collections are excluded from Select All while a search is active. No action required.Audio Notes: Capture spoken notes from the Pria toolbar. Tap the mic icon to open the Audio Notes modal, tap to start and stop (mobile-friendly toggle, no press-and-hold), and add as many segments as you need with playback, rename, and reorder. Hit Save and Pria transcribes in the background — notes appear in a new Audio Notes collection in your personal vault, each auto-titled from what you said. An optional Combine audio into a single note checkbox merges a session of short bursts into one note. Notes are linked to the active assistant and conversation, and token usage counts against your normal credit balance. Enabled by default for personal accounts and new twins; admins can disable it under Personalization → UI & Appearance → Disable Audio Notes for Users (also hidden in Guest UI or when file upload is disabled). Existing instances had it off at rollout — admins opt in explicitly.Large media & archive uploads flow through the queue: Long audio/video files that split into 10-minute chunks, and ZIP archives that expand into multiple documents, now appear as individual rows in the Files panel and process independently — each with its own live progress, retry budget, and final state. The parent row is hidden once its children are queued, and children expand automatically so you see exactly what was produced.Live file ingestion progress: Uploaded files now show live processing progress in the status tooltip with a percentage that climbs as the file is processed. The Files panel auto-polls in-flight uploads every few seconds, so you no longer need to refresh to see when a large PDF or video transcript is ready. Transient failures retry automatically; only terminal errors surface as “Processing failed”.File reprocess endpoint (API): New POST /api/user/files/{fileId}/reprocess re-queues an existing IP Vault file without re-downloading the source. Pass { "mode": "full" } to wipe all chunks and restart, or { "mode": "embed" } to keep chunks and only regenerate embeddings (useful after switching an embedding model). Returns 202 Accepted, or 409 if a job is still in flight. The existing Reprocess Content UI dialog works unchanged.Model context sizes reconciled: Audited every provider against official docs and corrected declared input/output token limits across Bedrock, OpenAI, Gemini, Mistral, and xAI. Highlights: us.amazon.nova-2-lite-v1:0 output 10K → 64K; gpt-5-mini / gpt-5-nano input 272K → 400K; codestral-2508 input 128K → 256K; us.meta.llama3-2-11b-instruct input 1M → 128K; Nova Pro/Micro output 10K → 5K; xAI Grok 4.1-fast now declares its 30K output cap. Metadata only — no behavior change, just an honest model picker.Rename conversation with AI: The Manage Conversation dialog now has a Rename with AI button that suggests a short, descriptive title from the first and last few dialogues — mirroring Summarize with AI. Review the suggested name in the editable field and save with Update.Unified RAG segment size: The admin RAG Segments editor now uses a consistent 8,000-character soft cap across all embedding models. The new Ollama option mxbai-embed-large (1024-dim, 512-token context) auto-scales ingestion to safer smaller segments, and what gets indexed now matches what you can edit.Vector dimensions in the model picker: The embedding model picker now shows Vector Dim (e.g. 1536d, 3072d) in place of Output Size, making models easier to compare. Two new models were added: mxbai-embed-large (Ollama, 1024-dim) and gemini-embedding-2-preview (Google, 3072-dim, with a raised 8,192-token input limit).Friendlier custom-LLM auth errors: When an institution’s custom AI model (Bring-Your-Own-Token endpoint) rejects a request with a 401/403, users now see a clear “your administrator needs to verify the API key in Settings → AI Models” message instead of the raw upstream error. Diagnostic emails to admins include the offending model’s name, URL, and provider so misconfigurations are easy to identify./api/user/tools minimum response: The runtime tools endpoint now accepts {"minimum": true} to omit the verbose instructions field from each tool — handy for selector UIs and BYOT integrations that only need the identifier and short description. Default behavior is unchanged.Tool-use uploads pinned to personal vault: Files saved via AI tools (generated images/audio/video, Google Drive downloads, URL scrapes) now always land in a standard user’s personal vault. Admins keep the existing behavior and can still route tool outputs to instance or account vaults.Cross-institution collection access: Fixed an issue where authorized admins got a 403 when opening collections (or adding/removing files) that belonged to another institution. The check now matches the entitlement-based behavior used elsewhere in the Files panel.Files panel stuck loading: Fixed a case where the Files panel stayed on “Loading, please wait…” after a failed collection fetch. It now falls back to the empty state so the error banner is visible and the panel stays usable.Embedded Files panel scroll loop: Fixed an infinite height growth when opening a collection from the admin Instances → IP Vault panel. The embedded Files container now uses a bounded viewport height.Claude Opus 4.7: Added support for Anthropic Claude Opus 4.7 on Bedrock (global.anthropic.claude-opus-4-7, us.anthropic.claude-opus-4-7) and the Anthropic Direct API (claude-opus-4-7). Opus 4.7 uses adaptive thinking with a native 1M input / 128K output context window (no extended-context toggle required). Opus 4.6 remains available and fully supported.Files entitlement system: IP Vault access is now governed by granular files.list, files.add, files.edit, and files.delete entitlements, so account admins can scope exactly who manages instance and account vaults. Standard users can no longer upload to instance vaults. Existing admin memberships are seeded with the new entitlements automatically — no action required. The Files panel shows only the actions each admin is entitled to.Personal upload gate enforced server-side: The institution setting Disable File Upload for Users is now enforced on the server — standard users get a 403 when uploading personal files (via file picker or URL) while it’s enabled. Allowed Email Domains for File Upload still carves out per-domain exceptions. Collections and existing files stay editable; the gate applies to new uploads only. The Files panel shows a Your access pill in Vault Health reflecting your entitlement when switching between personal, instance, and account vaults.Excluded Only file filter: The Files panel now has an Excluded Only toggle alongside Included Only, to quickly audit files not part of the RAG selection. The two are mutually exclusive; toggling a file’s inclusion while a filter is active removes it from the view.Video generation with Nova Reel and Sora 2: Video generation now routes through Amazon Nova Reel 1.1 (Bedrock, default) or OpenAI Sora 2 / Sora 2 Pro — configurable per institution from admin settings. Nova Reel makes 6-second clips from text or an image; Sora 2 makes 4, 8, or 12-second clips from text (quality='high' selects Sora 2 Pro). The generate_video tool works in chat and convo modes — source image optional, aspect ratio (16:9 / 9:16 / 1:1) defaults to 16:9, and the MP4 saves to your IP Vault. Stability AI’s video endpoint was retired by the provider, so stability.image-to-video is now deprecated.Stability AI native provider: Stability AI is now a first-class provider (alongside OpenAI, Anthropic, Google, Mistral, xAI, and Bedrock) for image and audio generation. It introduces the generate_audio tool (chat and convo modes) alongside generate_image. Stable Audio 2 generates up to 190 seconds of audio from text, saved to your vault. Admins can pick per-category models (Image / Audio Generation) and override the API key per institution.Bedrock legacy model deprecations: Marked us.amazon.nova-premier-v1:0, cohere.command-r-plus-v1:0, and amazon.nova-canvas-v1:0 as deprecated after AWS flagged them Legacy, and added cohere.command-r-v1:0 as the active Cohere-on-Bedrock replacement. Institutions on the deprecated conversation models should switch to us.amazon.nova-pro-v1:0 / us.amazon.nova-2-lite-v1:0 or cohere.command-r-v1:0; for image generation, switch to a Stability or OpenAI model.Cross-twin conversation search: From the gallery, search your conversation history across all your active digital twins. Each result shows the matching twin with shortcuts to View the record or Go To the conversation directly — switching twins automatically when needed.Login security hardening: Strengthened security across every sign-in path — email/password, Google, GitHub, Facebook, custom OAuth2 SSO, and Pria SDK launches — protecting SSO logins from cross-site forgery, closing a password-validation edge case, and keeping one-time login codes out of browser history, server logs, and referrer headers. Existing sessions keep working; no action required.Fix max-tokens dispatch: All providers (OpenAI, Anthropic, Google, Bedrock, xAI, Mistral) now correctly honor the institution’s “Max Tokens (Completion)” setting; previously xAI and Mistral ignored it and OpenAI defaulted to 32,000. “Unspecified (LLM Default)” now uses each provider’s native default, and “Auto” computes from the model catalog’s output cap. Bedrock thinking-mode users on “Auto” will see longer thinking budgets — pick an explicit value ≤16,000 if you hit throttling. “Unspecified” combined with thinking mode isn’t supported on Bedrock — use “Auto” or an explicit value ≥10,000.Admin edit modals: Every admin edit page (Instances, Accounts, Users, Assistants, AI Models, SSO, Tools, Questions, Feedbacks, Sessions, Histories, Saved States) now opens as a Lightbox over its list page, so you keep your scroll position, filters, and pagination. Prev/Next chevrons (and Ctrl+Arrow shortcuts) move between rows without closing the modal. Delete confirmations are now consistent Lightbox dialogs with an entity summary card. External deep-links (e.g. Canvas LMS SDK, Pria.js) continue to work.Institution lifecycle management: The admin Instances list now helps manage institution lifecycles. A status filter (Active | Inactive | All) switches views, and a Last Activity column shows a traffic-light dot (green within 30 days, yellow 30–90, red over 90). The Manage dialog offers Mark as Inactive on a stale active instance or Reactivate on an inactive one, and long-idle inactive instances (over 90 days) show a Ready for removal badge next to the Delete button.Inline account switcher: Admins with access to multiple accounts and the institutions.edit entitlement can reassign an instance to a different account directly from the Instances list — click the pencil next to the account name, pick a target, and confirm. The row updates in place without reloading.Last Activity on Accounts and Users lists: The admin Accounts and Users lists now show the same Last Activity column and stale indicator as the Instances list — traffic-light dots, relative time since last activity, and a yellow “inactive” hint for rows idle 30+ days — helping spot dormant accounts and users.Per-row entitlement scoping: Fixed an issue where an admin with membership in one institution saw Edit, Manage, and Delete on all rows in the Instances list — even institutions where they only had plain membership. Entitlement checks are now scoped per row, so admin actions only appear where you actually have permission.Sidebar UX redesign: Redesigned the Pria sidebar with a dedicated Conversation Model label and an entitlement-gated Configure button for jumping to model settings. Digital Twin and TTS Audio sections use a clean card layout, and new About RAG Modes and About Document Search help sections were added. The Tools section is now an always-visible scrollable list showing the enabled count. The “Load Max Dialogues” button was removed — infinite scroll handles it.Welcome screen enhancements: The Pria welcome screen now has flip cards that reveal descriptions on hover, a 3-column layout, and a subtle shimmer on the “What can I help with?” title. The streaming dialog uses a consistent input-box width and a colorful bouncing loader.Streaming scroll UX: A smoother streaming experience — your input and loader scroll into view immediately on submit, the completed dialog is fetched before clearing the stream to prevent jumps, and the view settles smoothly on the completed message.Admin Lightbox Manage modal: The Instances list Manage action is now a two-tier Lightbox (main + nested Confirm) instead of a popup. Release Credits is hidden when an instance has no parent account, and action buttons match their icon colors (amber for Recover, orange for Delete History). All admin list pages use a consistent action-menu wrapper.Base64 image support on custom LLMs: Custom OpenAI-compatible LLMs (Bring-Your-Own-Token) now accept base64-encoded images in the conversation, matching the native OpenAI provider — useful for vision-capable custom endpoints that don’t support URL image input.IP Vault UX improvements: Owners can now include/exclude, download, and move account-shared files regardless of their current institution. The Status badge (Included/Excluded) shows on all account-shared files, included files get a blue border, collection lists are deduplicated, and the Source field shows the assistant name with the instance in parentheses for clearer attribution.PDF OCR memory optimization: Reduced memory use during PDF OCR ingestion — page concurrency is capped, rendering scale was lowered, and file I/O is now async — preventing out-of-memory crashes on large image-based PDFs.Anthropic prompt cache limit: The Anthropic provider now caps prompt cache breakpoints at the API maximum of 4, preventing “too many cache_control blocks” errors on long conversations with system prompts.Android Chrome login fix: Fixed a blank blue page that some Android Chrome users saw after Google sign-in.Empty stream guard: Added empty-stream protection across all six AI providers (Bedrock, OpenAI, Anthropic, Google, Mistral, xAI) to prevent “No Content” errors when an upstream API returns an empty streaming response.Files panel stability: Fixed an issue where the Files panel could close unexpectedly when clicking the backdrop during upload, file move, or ingestion.Admin Instance About field: The About field in the admin Instance edit form is now a code editor with markdown syntax highlighting and a character counter, making long instance descriptions easier to write and proofread.Assistant RAG collections: Assistants can now be scoped to specific document collections in the IP Vault — the assistant’s retrieval only searches the selected collections, enabling focused assistants that reference specific knowledge bases while sharing the same IP Vault.Per-assistant tool configuration: Administrators can now control which AI tools each assistant can use. Set an assistant to “Selected Tools Only” and pick specific tools from the institution’s enabled set, with a filter toggle to view only selected tools or expand to the full list.Assistant-scoped RAG search: The standalone RAG Search tool and the MCP search-instance-rag endpoint now accept an optional assistant ID to scope results to that assistant’s configured collections. Results show collection names and the number of files searched, even when there are no matches.Faster RAG retrieval: Vector search across your vault is now significantly faster at scale.PDF OCR fix: Fixed an issue where image-based PDF pages failed OCR because the original PDF was sent to the vision step instead of the converted page image.xAI Grok native provider: Added xAI as a first-class AI provider with native SDK integration — streaming conversations, tool calling, vision, TTS (via ElevenLabs bridge), STT, and embeddings now route through xAI’s API directly. Supports Grok 3, Grok 3 Mini, and Grok 2 Vision models.Updated AI models: Added GPT-5.4, xAI Grok-4.20, Stability AI image generation models, and new embedding models to the platform catalog. Gemini 3.1 Flash Live Preview is now the default for Google real-time conversations.Admin Personalization reorganization: Reorganized the Personalization tab with a dedicated Convo (Real-Time Speech-to-Speech) section, improved Document Summarization placement, and clearer sub-section visual hierarchy.Voice preview links: Added direct preview links to the OpenAI TTS Playground and Google AI Studio so you can audition a voice before selecting it.ElevenLabs credential warning: A warning pill now appears in the Personalization tab when the ElevenLabs Agent ID or API Key is missing, guiding admins to the Configuration tab.Lightbox stability: A new preventBackdropClose option for modal components prevents accidental dismissal during critical workflows.Canvas course filtering: Fixed an issue where conversation history wasn’t filtered by course context when launching Pria via the Canvas LMS SDK.Canvas section reorder: The Canvas integration section moved above Voice Cloning in the Configuration tab for better discoverability.SDK iframe compatibility: Added localStorage and cookie shims for sandboxed iframes, so Pria works reliably when embedded in third-party LMS platforms with restrictive sandbox policies.Voice media embeds: Fixed an issue where media embed URLs in AI responses were corrupted by the voice sanitizer, preventing audio and video previews from rendering correctly.Documentation audit: Comprehensive update of all documentation to reflect current model catalogs, tool counts, voice providers, and admin UI layouts.RAG mode selector: Choose between three RAG modes — Normal (RAG-augmented responses), Search Only (returns only RAG results without AI generation), and Disabled (pure LLM responses). Available in the Pria sidebar, as a toolbar cycle button, and in your profile page.Admin Histories pagination & filtering: The admin Histories page now uses server-side pagination with infinite scroll instead of loading all records at once. Added a Course Name filter, and the charts (Model Distribution, Peak Usage Heatmap) now respect all toolbar filters including User, Account Type, Course, and Date Range.Admin server-side pagination: All 14 admin list pages now use server-side pagination with infinite scroll instead of loading all records at once — Users, Entitlements, Feedbacks, Institutions, Sessions, Accounts, AI Models, Assistants, Tools, SSO Providers, Payments, Saved States, Questions, and System Sessions. A “Showing X of Y” record count appears above each table, and additional pages load automatically as you scroll.- Mistral AI Native Provider: Added Mistral AI as a first-class provider — conversation, vision, reasoning, code, multilingual, TTS, STT, embeddings, and moderation models now run natively for better performance and full feature support
- YouTube URL Import: Paste a YouTube URL into the uploader and Pria automatically pulls the transcript with timestamps, video metadata, and thumbnail. Videos without captions can still be imported with Skip Indexing enabled, and re-ingesting refreshes the transcript and thumbnail
- Reprocess Content: Redesigned Reprocess Content dialog with dedicated cards for re-embedding and re-ingesting, a Skip Indexing option, and clearer error handling
- Skip Indexing: Upload files, URLs, or clipboard content to your IP Vault without indexing them for AI search — useful for storage-only files or content you want to index later. A per-file “Skip Indexing” checkbox is available on all upload tabs, and skipped files can be reprocessed anytime
- SSO Provider Management: Admins can configure OAuth 2.0 Single Sign-On providers from the Admin dashboard — set identity provider endpoints, client credentials, and field mappings, then assign them to institutions. Users sign in at
/sso/{slug}and are automatically logged in and enrolled, with connectivity testing and dark mode support on the SSO login page - File Preview: Enhanced fullscreen file preview with HTML document rendering (Preview and Source views), print support for all file types, and an open-in-new-tab button
- Guest Mode: Institutions can enable a minimal guest interface that hides the sidebar, file uploads, assistant switching, and profile menu — ideal for demos or guided experiences
- Lightbox Wizards: Create and Enroll Digital Twin flows now open as overlay modals above the Gallery, so you never lose your place while browsing
- Copy Feedback: Clipboard copy actions now show a smooth 3D icon-flip animation instead of popup alerts, with clean print output in both light and dark mode
- Gallery Favorites: Toggle favorite Digital Twins directly from the Gallery cards for quick access
- ElevenLabs Transport: Admins can choose the connection method for ElevenLabs voice conversations
- RAG Search Timing: Search results now show how long the search took, for performance transparency
- Google Document Upload: Paste Google Docs, Sheets, or Slides URLs directly into the uploader — content is read securely via Google OAuth. Supports personal and institution-level Google authorization, with context-aware prompts to connect your account when needed
- Google Slides Reader: Google Presentations are now fully supported in URL uploads, extracting text from all slides
- RAG Segments Panel: Improved segment management with metadata editing and a one-click reprocess workflow
- Gemini Live Stability: Improved reconnection handling for Gemini voice conversations, with automatic retry on unexpected disconnects
- Smarter Conversation Memory: Pria now compacts older conversation history into concise summaries while keeping your most recent exchanges verbatim, so longer conversations stay coherent and responsive without burning extra credits. Activates automatically when Remember History is set to 6 or higher
- Streaming Performance: Dramatically smoother AI response streaming, eliminating visible lag during long responses
- Typing Responsiveness: Fixed input lag during AI streaming — typing in the text entry bar is now instant even while a response is being generated
- Stripe Billing: Stripe-powered credit purchasing with personal and institutional packages, purchase confirmation screens, and a new admin Payments page with full payment history
- Conversation Manager: Manage conversations and favorites in Lightbox popups for a cleaner, more focused experience
- Mini Sidebar: Redesigned sidebar with hover expansion and instant search-as-you-type for quick access to conversations and tools
- Options Tray Redesign: Animated collapsible options panel with a soft glow border for a more polished settings experience
- Creation Wizard: Rebuilt the Digital Twin creation flow as an intuitive 8-step wizard with step animations
- PDF Generation: AI can now generate formatted PDF documents via the
save_to_filetool, with math symbols and special character support - Drag & Drop Upload: Upload files by dragging them onto the interface or pasting from clipboard, with a new fullscreen file preview
- Code Editor: Rich code editor with live markdown preview for editing instructions, replacing plain text areas
- Summary View: Redesigned summary display with a loader animation and streamlined follow-up UI
- Multi-Provider Embeddings: Document search embeddings now support Bedrock Titan, Google Gemini, and OpenAI models, selectable per institution
- RAG Chunk Enrichment: Smarter document processing for more accurate search results
- Embedding Segments: New UI for creating and managing individual embedding segments
- OpenAI Realtime Enhancements: Noise reduction, configurable transcription language, and improved session defaults for Convo Mode
- Gemini Realtime Improvements: Fixed transcript accumulation and voice separation in Gemini-powered conversations
- Download Tracking: Upload view now surfaces download statistics and file usage analytics
- Vault Health: File integrity indicators visible for all users across all file statuses
- iPhone & QuickTime Audio Support: Upload
.movrecordings directly — Pria automatically extracts and converts the audio before transcription, so iPhone voice memos and QuickTime exports work seamlessly across all AI providers - API Key Authentication: Admins can now authenticate via API key for programmatic access — generate, view, copy, and revoke personal API keys directly from the Profile page
- Admin Histories Filter: New account type dropdown in the Histories admin page lets you filter conversation records by user or admin accounts, with filter state preserved across pagination
- Clipboard Restrictions: Institutions can disable copy/paste for non-admin users to protect sensitive AI-generated content, while admins retain full access
- Animated Loading Screen: Redesigned loading experience with an animated Pria logo, radial fade, and smooth fade-in transitions
- Bundle Optimization: Faster page loads through code splitting
- Security: Inactive or pending institution members are blocked from AI interactions; all active sessions are revoked on password change; Facebook login moved to server-side OAuth2; password policy enforcement; SVG sanitization; and vulnerability remediation
- Dark Mode: Fixed select dropdown and input background inconsistencies in dark mode
- File Downloads: Corrected file extension mismatches in downloaded files
- Deep Research: Reasoning effort automatically tuned for deep-research model configurations
- Uploads Access Control: Three access levels for files — owner, institution members, and shared
- iOS Safari: Fixed keyboard toolbar overlapping the chat input area on iOS Safari
- Gemini Flash Live: Added Google Gemini 2.5 Flash as a third real-time voice provider in Convo Mode, alongside OpenAI Realtime and ElevenLabs, with tool calling, transcription, and per-institution voice selection
- Claude Sonnet 4.6: Added support for Claude Sonnet 4.6 and set it as the new platform default model
- Gemini 3.1 Pro: Added Gemini 3.1 Pro to the model catalog
- GPT Realtime 1.5: Added OpenAI gpt-realtime-1.5 and set it as the default for Convo Mode
- Direct AI Providers: Introduced native Anthropic and Google GenAI providers for direct model access, improving performance and enabling full feature support (tool use, streaming, prompt caching)
- Reasoning Effort: Added configurable thinking/reasoning effort levels alongside max token settings for supported models
- Custom AI Models: Added validation for custom model configurations (API URL, API key) to prevent misconfiguration, plus deprecated-model indicators in the catalog
- File Collections: Organize files into nested collections with drag-and-drop, breadcrumb navigation, and animated transitions, plus a bulk edit mode for multi-select move, delete, and reprocess
- Files Redesign: Rebuilt the Files interface with a fast grid, infinite scroll, live upload progress indicators, and enhanced file cards with type-specific icons
- Upload Experience: Redesigned upload workflow with deferred processing, RAG segment editing, and Google Drive deduplication
- Chunked Audio/Video Transcription: Large audio and video files (>20MB) are automatically transcribed with speaker diarization support, raising the upload limit to 150MB
- Excel Legacy Support: Added support for older
.xlsExcel spreadsheets - History Compaction: Conversation history compaction keeps the 3 most recent exchanges verbatim while summarizing older context, reducing token usage in long conversations. Configurable per institution
- Extended Context: Added a toggle to enable 1M token extended context for supported models, shown only when the selected model supports it
- Vault Health: New health panel in the Files vault showing error counts, unused files, and content quality scores
- RAG Error Recovery: Files stuck in processing are now automatically marked as failed with a red indicator, and stale uploads (>15 minutes) offer one-click Delete or Reprocess
- Edit Assistant Redesign: Redesigned the Edit Assistant lightbox with improved variable management and dynamic variable mapping
- ElevenLabs Voice: Added ElevenLabs as a Realtime speech-to-speech voice provider with voice cloning support in Digital Twin personalization
- Google Tools in Convo Mode: Google services (Drive, Calendar, Gmail) are now accessible during voice conversations
- Session Analytics: New charts in the admin Sessions and Histories tabs — device type distribution, activity timelines, peak usage heatmaps, and model distribution with selectable chart types
- Google OAuth Multi-Institution: Added context-aware Google OAuth supporting authentication across multiple institutions with automatic token validation
- Account Manager Role: Introduced a new Account Manager role for managing multiple institutions within an account (contact humans@praxis-ai.com to request access)
- Enroll Users: Added an Enroll Users tab to the admin Entitlements page for streamlined onboarding with auto-loading of eligible users
- Platform Monitoring: Enhanced platform-level session monitoring with per-transport hourly stats and persistent client session history
- Admin Filtering: Added an AccountSelect filter across all admin list pages for easier multi-account management
- Security Hardening: Signed SDK launch tokens, field whitelisting on all update endpoints, hashed reset codes, rate limiting on auth endpoints, and privilege escalation fixes
- Dark Mode: Extended dark mode to loading dialogs, info banners, branding backgrounds, login buttons, and Canvas token cards across 18+ components
- Profile Auto-Save: Profile setting toggles now auto-save on change without a manual save action
- Feedback Redesign: Redesigned the feedback interface as a Lightbox modal for a more focused experience
- Touch Support: Added touch event support for improved mobile and tablet interactions
- Enrollment Search: Enhanced the enrollment page with keyword search
- Performance: Optimized file queries, assistant loading, connection handling, and session charts for faster page loads and fewer admin timeouts
- Dark Mode (Beta): Released
darkmode support in the Pria UI, with custom dark background images per institution - Digital Twin Gallery: Launched a browsable gallery for discovering and selecting Digital Twins, with ElevenLabs voice embedding support
- Text Entry Repositioning: Resized and repositioned the text entry bar to float over the conversation history for better visual hierarchy and more screen space
- Files and Assistants Layout: Repositioned Files and Assistants under the text entry bar for a more compact, streamlined display, with redesigned file and assistant cards
- Profile Picture Upload: Added the ability to upload profile pictures from webcam or file
- Sidebar Enhancements: Collapsible sidebar tray, direct tool enable/disable toggle, and a RAG-specific configuration panel
- Terminology Update: Renamed “Agents” to “Tools” across the platform; instances are now consistently called “Digital Twins”
- Institution Cloning: Administrators can now clone an entire institution configuration
- Streaming API: Added a Q&A streaming endpoint for real-time response delivery
- Prompt Optimization for 2026 Models: Tuned system instructions for the latest models (GPT-5.2, Claude 4.5, Gemini 3.0 Flash), prioritizing personality-first responses for more immersive Digital Twin experiences
- Tool Guidance Toggle: Added a switch to reduce instruction verbosity for general conversations while keeping full guidance for advanced research queries, improving response speed without sacrificing quality
- Model-Aware Response Tuning: Tuned responses to leverage the latest models’ capabilities — improved reasoning, fewer hallucinations, and better context retention across long conversations
- Voice Selection: Enhanced voice picker for Convo mode with an improved selection experience
- SDK Hardening: Improved SDK reliability with inactive-instance handling and membership validation
- File Cleanup: Added an automatic cleanup routine to remove orphaned files from deleted users and institutions
- Open AI GPT 5.2: Added support for the latest GPT 5.2 model from OpenAI.
- Manage Favorites: Rename and organize your favorites with custom descriptions or AI-generated summaries.
- GitHub Login: Added support for authentication via GitHub.
- Google Login: Enhanced authentication process for Google accounts.
- Account-Level Assistants: Assistants can now be shared at the account scope, making them instantly available to all digital twins within the same account.
- Account-Level RAG Files: Documents can now be promoted from Instance to Account level, giving every digital twin in the account access to a shared knowledge base.
- History Graph: Now defaults to the last 12 weeks with daily data shown as a bar chart.
- Assistant Filtering: Added a toggle to quickly filter assistants by type: Personal, Instance, or Account.
- Model Tracking: Automatically records which AI model was used for content summarization and image analysis when uploading files and URLs.
- Sonnet 4.5: Added support for the Claude Sonnet 4.5 model and set it as the platform default.
- All Anthropic Models: Upgraded plans to support large input context up to 1 million tokens.
- Tool Upgrade: Improved
call_canvasandsearch_canvasto run simultaneously. - Content Length Returned by Tools: Tools now return up to 60,000 characters (increased from 20,000).
- Assistants Features: New options to disable RAG searches when running assistants, and to run assistant instructions without system context.
- Pria JS SDK: Added a
pria.isReady()function that returns whether Pria is ready to receive messages. - File Import Enhancements: Improved support for ZIP and IMSCC file imports.
- Shared Memory: Added shared memory that stores information accessible across all users of the same digital twin, enabling cross-user persistence and collaboration.
- Memory Tools Enhancement: Unified the memory tools (
set_in_memoryandget_from_memory) to support both user-specific and shared memory. - Privacy Policy Enforcement: Introduced account-level privacy policy settings that require users to accept terms before accessing Digital Twins.
- Selective File Upload Authorization: When file upload is disabled for users, administrators can authorize selected user groups by email domain suffix.
- Selective Assistant Authorization: When Assistants are disabled for users, administrators can authorize selected user groups by email domain suffix.
- Favorite Digital Twins Priority: Digital twins marked as favorites now appear first in the gallery carousel for easier discovery and quick access.
- OCR for Image-Based PDFs: Image-based PDFs (with no extractable text) now automatically undergo OCR for text extraction and indexing.
- Enhanced Date Range Selection: Administrators can now filter history and session records using a date range picker.
- Documentation Updates: Comprehensive updates covering RAG (Retrieval-Augmented Generation), Bring Your Own Token (BYOT), the
call_canvasandsearch_canvastools, and more at https://docs.praxis-ai.com
- Anthropic Sonnet 4.5: Support for Claude Sonnet 4.5 (20250929) with a 200K context window, 64K output tokens, and enhanced multimodal capabilities including vision, document analysis, and tool use.
- MCP Server: Introduced an MCP Server that lets any AI solution call your Digital Twin.
- Security & Performance: Improved RAG security and performance.
- REST APIs: Expanded functions in the Runtime and Admin REST APIs.
- Documentation: Major updates—especially for Assistants, Canvas integration, and MCP at https://docs.praxis-ai.com
- Convo Mode Input: Added text input for conversations in full screen; reply by text while receiving vocal responses.
- Transcripts in Full Screen: Live transcripts available during Convo mode in full screen.
- Password Reset: Users can be required to reset their password on next login.
- Logged Remote Calls: Remote connector (MCP) calls are now logged in Tools Details.
- Gepetto Digital Twin: Improved AI Portrait Artist for generating Digital Twin avatars.
- Moderation by Default: Moderation is now enabled by default when creating new Digital Twins.
- Connector MCP: Added support for Remote Model Context Protocol (MCP) servers, individually configurable for each Digital Twin instance.
- Custom LLM Integration: Support for custom Large Language Models (LLMs) with per-Digital Twin configuration, letting organizations deploy specialized models for specific use cases (e.g. Vertex).
- Digital Twin Gallery Enhancements: Improved navigation with intuitive arrows and resolved display flickering on macOS for smoother interactions.
- OAuth Verification for Canvas Integration: Added mandatory OAuth 2.0 authentication for Canvas Theme SDK access, ensuring secure user verification before Digital Twin initialization.
- Streamlined Digital Twin Creation: A simplified 2-step Digital Twin creation process directly through the Gallery, reducing deployment complexity and time-to-value.
- Support for Mermaid diagrams and charts: Added support for Mermaid Diagramming and Charting.
- User Experience: Enhanced digital twin loading for standard users.
- Branding: Added branding for personalized login/signup.
- Text Customization: Customizable activation and joining messages.
- Logo Branding: Added a logo option for login sequence customization.
- Animation: Added animated picture support for Digital Twins.
- Token Management: Improved Canvas faculty token refresh.
- Digital Twins: Added a gallery for twin selection and management.
- REST Tools: Custom REST-enabled tool creation for instances.
- Email Notifications: Enhanced approval workflow notifications.
- Speech Features: Retired a deprecated speech component.
- Accessibility: Improved contrast ratios and ARIA labels.
- Substitution Variables: New prompt variable management system.
- Email Tool: New email functionality.
- Conversation Persistence: Added save/restore for your last conversation.
- Search Enhancement: Integrated Exa.ai for semantic web search.
- xAI Integration: Added support for xAI models.
- SDK Toolkit: Enhanced UI development capabilities.
- File Upload: Expanded support for all file types across platforms.
- Credit Transfers: Enhanced account-to-instance credit management.
- File Tools: Added a get_user_files tool for document collection access.
- Instance Management: New comprehensive admin management system.
- Performance: Improved file loading with asynchronous summaries.
- Pricing: Decreased costs across personal accounts.
- Email Processing: Improved email content parsing.
- Platform Upgrade: Backend updates for improved compatibility and stability.
- Streaming: Enhanced conversation streaming.
- Canvas Scoping: Added course ID filtering for history records.
- Assistant Management: Made user assistants admin-visible for sharing.
- AI Models: Added Claude 3.7 Sonnet support.
- Conversation Mode: Extended user access to conversation features.
- File Management: Enhanced file handling and RAG integration.
- Mistral Integration: Added direct API access for Mistral models.
- Bug Fixes: Resolved styling and file system errors.
- Accessibility: CSS improvements for users with epilepsy.
- Canvas SDK: Updated to version 1.3 with enhanced user ID handling.
- Security: Addressed third-party component vulnerabilities.
- Navigation: Improved UI layout for large screens.
- API Integration: Enhanced Canvas API access with the call_canvas tool.
- SDK Features: Added browser messaging capabilities and documentation.
- Data Compliance: Added DPA compliance features for data deletion and recovery.
- Canvas Integration: Added a discussion bot plugin.
- Token Management: Enhanced Canvas token refresh.
- Database Improvements: Improved database performance.
- Reliability: Improved retry handling for model requests.
- Analytics: Added tool usage tracking and metrics.
- Instance Management: Added the ability to disable specific assistants per instance.
- Email Notifications: Added low-credit alerts for administrators.
- Location Services: Enhanced web search with location-based results.
- Document Summarization: Improved summarization quality and reliability.
- Course Management: Added courses in the sidebar for institutional scoping.
- Canvas tool: Enhanced interpretation and display capabilities.
- Student Analytics: New feature to analyze student questions.
- Assistant Features: Launched new Assistant functionality and a Course Syllabus Generator.
- Bug Fixes: Resolved input size limitations and tool usage issues.
- Admin Enhancements: Multi-select instance support and Gemini Flash 1.5.
v0.0.10
- LTI Brightspace: Bug fixes for course parameter passing.
- SSO Integration: Implemented QuickLaunch SSO OAuth.
- Credit Management: Added admin ability to transfer credits between accounts.
- AI Models: Added support for o1-preview and o1-mini models.
- Scalability: Fixed entitlements listing performance issues.
- Scientific Literature: Added a CrossRef tool for academic research support.
- Real-time Communication: Implemented OpenAI WebRTC APIs with UI.
- Voice Features: Enhanced voice selection capabilities.
- Canvas Integration: Captured global ID for Canvas OAuth tokens.
- User ID Updates: Relaxed user ID requirements to support string format.
- UI Improvements: Various interface enhancements and bug fixes.
- UI Enhancements: Implemented anchor text conversion for links.
- Copy Functionality: Added copy-input-to-clipboard feature.
- Timeout Resolution: Fixed response timeout issues.
- SDK Integration: Developed embed SDK for LMS integration.
- Personality Customization: Added ability to swap AI personalities per institution.
- Code Block Support: Enhanced parsing for multiple code blocks.
- Image Display: Added HTML image tag support.