What's new? - praxis-ai

April 2026

Display Thinking Details (admin toggle): Pria now collects model reasoning / chain-of-thought from the providers that expose it (Anthropic, AWS Bedrock, Google Gemini, OpenAI summaries, xAI) and renders it in a collapsible 💡 lightbulb block above the assistant response. Mistral does not expose thinking and shows no lightbulb. The new Display Thinking Details checkbox on the institution Personalization page (next to Display Tools Details) controls whether thinking is streamed, displayed, and persisted to History — defaults to on. Personal users always see thinking; institution-scoped accounts inherit the toggle. Streaming consumers receive a new THINKING event type on Socket.IO RECEIVE_STREAM and the SSE stream — see the API Reference / Streaming docs for the payload shape (per-round id/round/delta/text/done/signature). History list endpoints now expose only a hasThinking: boolean flag; the full thinking body is fetched on demand via GET /api/user/history/{id}/thinking (and the admin counterpart) when the user opens the lightbulb block — keeping list responses lean. Bundled regression fix: institutions that pre-dated the schema default for displayAgentDetails and rtAdminOnly no longer surprise-default to the wrong toggle value.Safer Bulk Delete in IP Vault: The Select All button in the Files panel now respects whatever filter you have active — searching for scr_* and clicking Select All picks only the matching files, not every file in the vault. The bulk-delete confirmation also lists every collection and file that will be removed in a scrollable preview so you can review before clicking the red button. Collections are excluded from Select All while a search is active (selecting a folder would have pulled in non-matching files inside it). No action required.Audio Notes: Capture spoken notes directly from the Pria toolbar. Tap the mic icon to open the Audio Notes modal, tap the microphone to start and stop recording (mobile-friendly toggle, no press-and-hold), then add as many segments as you need with playback, rename, and reorder controls. Hit Save and Pria transcribes the audio in the background through the existing ingestion pipeline — files appear in a new Audio Notes collection in your personal vault. Once transcription completes, each note’s filename is automatically replaced with a short LLM-generated title derived from what you said (e.g. Quarterly planning notes.webm instead of a raw timestamp). The save dialog stays open with a green confirmation and a Close button so you know the recording is on its way to processing. Optional Combine audio into a single note checkbox concatenates segments losslessly via ffmpeg before transcription so a session of short bursts is filed as one note instead of many. Notes are linked to the active assistant and conversation context just like file uploads, and token usage is accounted against your normal credit balance. Audio Notes is enabled by default for personal accounts and for new digital twin instances; instance administrators can disable it from Personalization → UI & Appearance → Disable Audio Notes for Users, and it is automatically hidden in Guest UI mode or whenever Disable File Upload for Users is on. To avoid surprising admins of existing instances, every pre-existing institution had Audio Notes turned off during the rollout — admins must explicitly opt their instance back in.Large Media & Archive Uploads Flow Through the Queue: Long audio/video files that get split into 10-minute chunks, and ZIP archives that expand into multiple documents, now appear as individual rows in the Files panel and progress through the ingestion queue independently — each with its own live progress, retry budget, and terminal state. Previously the parent file stayed “Processing…” while the children were embedded inline, which could pin the event loop and delay other uploads; now every resulting document is embedded as a queued background job. The parent row is hidden from the Files grid once all children are queued, and the children are expanded automatically after each poll so you see exactly what was produced.Live File Ingestion Progress: Uploaded files now show live processing progress in the status tooltip as they flow through the IP Vault pipeline — extract, chunk, sanitize, embed, finalize — with a percentage indicator that climbs smoothly as each chunk is embedded. The Files panel also auto-polls in-flight uploads every few seconds while they process, so you no longer need to manually refresh to see when a large PDF or video transcript is ready for use in RAG. Transient failures are retried automatically with exponential backoff; only terminal errors surface as the “Processing failed” state.File Reprocess Endpoint (API): New POST /api/user/files/{fileId}/reprocess endpoint re-queues an existing IP Vault file through the ingestion pipeline without re-downloading the source. Accepts { "mode": "full" } to wipe all chunks and restart from scratch, or { "mode": "embed" } to keep existing chunks and only regenerate embeddings — useful after switching an institution’s embedding model. Returns 202 Accepted with the queued phase; returns 409 if an earlier job is still in flight. The existing Reprocess Content UI dialog in the Files panel continues to work unchanged.Model Context Sizes Reconciled: Audited all providers against their official documentation and corrected declared input/output token limits across Bedrock, OpenAI, Gemini, Mistral, and xAI catalogs. Highlights: us.amazon.nova-2-lite-v1:0 output bumped from 10K → 64K tokens (previously under-declared 6×); gpt-5-mini and gpt-5-nano input raised from 272K → 400K; codestral-2508 input raised from 128K → 256K; mistral.devstral-2-123b and mistral.mistral-large-3-675b-instruct input tightened from 256K → 128K (matching AWS Bedrock model cards); us.meta.llama3-2-11b-instruct input corrected from 1M → 128K; Nova Pro/Micro output reduced from 10K → 5K; openai.gpt-oss-120b-1:0 realigned to 128K/16K; mistral-embed, codestral-embed, and Gemini Live models nudged to exact doc values. xAI Grok 4.1-fast now declares official 30K output cap; Grok Code Fast 1 declares 10K. No code behavior changes — these are metadata corrections that keep the model picker honest about each model’s true limits.Rename Conversation with AI: The Manage Conversation dialog in Pria now has a Rename with AI button that generates a short, descriptive title for the conversation from the first and last few dialogues — mirroring the existing Summarize with AI action on favorites. Click once, review the suggested name in the editable field, and save with Update.Unified RAG Segment Size: Admin RAG Segments editor now uses a consistent 8,000-character soft cap across all embedding models (instead of showing the underlying token limit interpreted as characters). The new Ollama option mxbai-embed-large (1024-dim, 512-token context) auto-scales ingestion to safer smaller segments, and all uploads benefit from the unified chunk-size constant shared between ingestion and the editor — no more divergence between what gets indexed and what you can edit.Vector Dimensions Surfaced in Model Picker: The embedding model picker now shows Vector Dim (e.g. 1536d, 3072d) in place of Output Size, making it easier to compare embedding models at a glance. The picker also adds support for two new models: mxbai-embed-large (Ollama, 1024-dim) and gemini-embedding-2-preview (Google, 3072-dim) — the latter with a raised 8,192-token input limit that allows larger RAG segments per chunk.Friendlier Custom-LLM Auth Errors: When an institution’s custom AI model (Bring-Your-Own-Token endpoints like LiteLLM proxies) 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 (which in some cases surfaced proxy internals and partial key fingerprints). Super-admin diagnostic emails are enriched with the offending custom model’s name, URL, and provider, plus the original upstream error — so misconfigurations can be identified without cross-referencing collections manually./api/user/tools Minimum Response: The runtime tools endpoint now accepts {"minimum": true} in the request body to omit the verbose instructions field from each tool, matching the shape of the admin tools endpoint. Useful for selector UIs and BYOT integrations that only need the identifier + short description. Default behavior is unchanged.Tool-Use Uploads Pinned to Personal Vault: Hardened the AI tool-use upload path so files saved via tools (generated images/audio/video, Google Drive downloads, URL scrapes triggered from a tool call) always land in a standard user’s personal vault. Admins keep the existing behavior and can still route tool-use outputs to instance or account vaults.Super Admin Collection Access: Fixed an issue where super admins received a 403 when opening collections (or adding/removing files) that belonged to another institution. The ownership and vault-scoping check in the collection file endpoints now bypasses for super admins, matching the entitlement-based behavior of the rest of the Files panel.Files Panel Stuck Loading: Fixed a case where the Files panel remained pinned on “Loading, please wait…” after a failed collection fetch. The error path now falls back to the empty state so the error banner is visible and the panel remains usable.Embedded Files Panel Scroll Loop: Fixed an infinite scroll-height growth when opening a collection from the admin Instances → IP Vault panel. The embedded Files container now uses a bounded viewport height so the grid’s self-measurement can no longer feed back through the resize observer.Claude Opus 4.7: Added support for Anthropic Claude Opus 4.7 across 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 the adaptive-thinking API 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 administrators can precisely scope who manages instance and account vaults. Standard users can no longer upload to instance vaults. Existing admin memberships are automatically seeded with the new entitlements on platform startup — no action required. The Files panel surfaces only the actions each admin is entitled to (for example, delete buttons are hidden from admins without files.delete).Personal Upload Gate Enforced Server-Side: The institution setting Disable File Upload for Users is now enforced on the server — standard users receive a 403 when uploading personal files (via file picker or URL) while the flag is enabled. Allowed Email Domains for File Upload still carves out per-domain exceptions, letting specific users retain personal-upload rights. Collections and existing personal files remain editable; the gate applies to new uploads only. The admin entitlement list now exposes the files.* module under Manage → Entitlements, and the Files panel shows a Your access pill row inside Vault Health that reflects your current 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, letting you quickly audit files that are not part of the RAG selection. The two filters are mutually exclusive; toggling a file’s inclusion state while a filter is active removes it from the view in place.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 generates 6-second clips from a text prompt or source image; Sora 2 generates 4, 8, or 12-second clips from text (quality='high' auto-selects Sora 2 Pro). The generate_video tool works in chat and convo modes — source image is optional, aspect ratio (16:9 / 9:16 / 1:1) defaults to 16:9, generated MP4 is saved to the user’s IP Vault. Stability AI’s video endpoint was retired by the provider on 2025-07-24 and the stability.image-to-video option is now marked deprecated (existing institutions pinned to it see a deprecation message — switch to Nova Reel or Sora 2).Stability AI Native Provider: Stability AI is now a first-class provider alongside OpenAI, Anthropic, Google, Mistral, xAI, and Bedrock — dedicated to Image and Audio generation. Introduces the generate_audio tool, available in both chat and convo modes, alongside the existing generate_image tool (generate_video now routes to Nova Reel / Sora 2 — see above). Stable Audio 2 generates up to 190 seconds of audio from text. Generated assets are saved to the user’s vault like any other upload. Admins can pick per-category models (Image / Audio Generation) and override the platform 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 as Legacy. Added cohere.command-r-v1:0 as the active Cohere-on-Bedrock replacement. Institutions pinned to 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 — Stability image models are no longer available through Bedrock and need to be used via the direct provider.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 — the twin switch happens automatically when needed.Login Security Hardening: Strengthened security across every sign-in path — standard email/password, Google, GitHub, Facebook, custom OAuth2 SSO, and Pria SDK launches. The changes protect SSO logins from cross-site forgery attacks, close a subtle password-validation edge case, and keep one-time login codes out of browser history, server logs, and referrer headers. Existing sessions continue to work; no action required.Fix Max Tokens Dispatch: All AI providers (OpenAI, Anthropic, Google, Bedrock, xAI, Mistral) now correctly honor the institution’s “Max Tokens (Completion)” setting. Previously, xAI and Mistral silently ignored the setting, and OpenAI defaulted to 32,000 when unset. The “Unspecified (LLM Default)” option now truly uses each provider’s native default (4,096 for Anthropic, 8,192 for Google, omit for the others), and “Auto” computes from the model catalog’s output cap. Bedrock + thinking-mode users who picked “Auto” will see significantly longer thinking budgets (the previous hardcoded 3,000-token cap is gone) — monitor for “Too many tokens” throttling and pick an explicit value ≤16,000 if affected. “Unspecified” combined with thinking mode is not supported on bedrock — pick “Auto” or an explicit value ≥10,000 to enable thinking.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 modal over its list page — you no longer lose scroll position, filters, or pagination when editing a record. Prev/Next chevrons (and Ctrl+Arrow shortcuts) let you navigate between rows without closing the modal. Delete confirmations across all admin list pages are now consistent Lightbox dialogs with an entity summary card instead of inline popovers. External deep-links (e.g. Canvas LMS SDK, Pria.js) continue to work — legacy /update/:id URLs are transparently rewritten to the new pattern.Institution Lifecycle Management: The admin Instances list now helps you manage the lifecycle of your institutions. A new status filter (Active | Inactive | All) lets you quickly switch between lifecycles, and a new Last Activity column shows a traffic-light dot (green for recent activity within 30 days, yellow for 30-90 days, red for over 90 days) next to each institution. Open the Manage dialog on a stale active instance to see a Mark as Inactive action, or on an inactive instance to see a Reactivate action. Long-idle inactive instances (over 90 days) also show a Ready for removal badge next to the Delete button, making it easy to spot candidates for cleanup.Inline Account Switcher: Admins with access to multiple accounts and the institutions.edit entitlement can now reassign an instance to a different account directly from the Instances list — click the pencil icon next to the account name, pick a target account from the dropdown, confirm the switch. The row updates in place without reloading the page.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 (green/yellow/red), relative time since last activity, and a yellow “inactive” hint next to the Status column for rows idle for 30+ days. Helps admins quickly spot accounts and users that have gone dormant.Per-Row Entitlement Scoping: Fixed an issue in the admin Instances list where an admin with membership in one institution would see Edit, Manage, and Delete actions on all rows — even institutions where they only had plain membership. Entitlement checks are now scoped per-row, so admin actions only appear on the institutions you actually have the permissions on.Sidebar UX Redesign: Redesigned the Pria sidebar with a dedicated Conversation Model label and entitlement-gated Configure button for quickly jumping to model settings. Digital Twin and TTS Audio sections now use a clean card layout. Added About RAG Modes and About Document Search expandable help sections with lightbulb tips for get_browser and tool configuration. The Tools section is now an always-visible scrollable list showing the enabled count. Removed the “Load Max Dialogues” button — infinite scroll handles it automatically.Welcome Screen Enhancements: The Pria welcome screen now features flip cards that reveal descriptions on hover, a 3-column grid layout, and a subtle shimmer animation on the “What can I help with?” title. The streaming output dialog uses a consistent input box width and a colorful gradient bouncing loader.Streaming Scroll UX: 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. Async callbacks are guarded against unmount to prevent stale state errors.Admin Lightbox Manage Modal: The admin Instances list Manage action is now a two-tier Lightbox modal (main + nested Confirm) instead of a popup. Release Credits is hidden when an instance has no parent account, action buttons match their icon colors (amber for Recover, orange for Delete History), and cancel handlers are hardened against missing events. All admin list pages now use a consistent ButtonBar wrapper for action menus.Base64 Image Support on Custom LLMs: Custom OpenAI-compatible LLMs (configured via Bring-Your-Own-Token) now accept base64-encoded images in the conversation — matching the behavior of the native OpenAI provider. Useful for vision-capable custom endpoints that don’t support URL-based image input.IP Vault UX Improvements: Owners can now include/exclude, download, and move account-shared files regardless of their current institution context. The Status badge (Included/Excluded) is visible on all account-shared files, and included FileCards display a blue border highlight. Collection lists in instance and account vaults are deduplicated to prevent root-level duplicates, and “Added By” metadata is batch-resolved for faster rendering. The Source field now shows the assistant name with the instance name in parentheses for clearer attribution.PDF OCR Memory Optimization: Reduced memory usage during PDF OCR ingestion — concurrency is now capped at 3 pages at a time (instead of unbounded), rendering scale dropped from 2× to 1.5× (~44% less memory per page), and sync file I/O was replaced with async. Prevents out-of-memory crashes on large image-based PDFs.Anthropic Prompt Cache Limit: The Anthropic provider now caps prompt cache breakpoints at 4 (the API maximum), counting existing cache points first to prevent “too many cache_control blocks” errors on long conversations with system prompts.Android Chrome Login Fix: Fixed a blue-screen issue where Android Chrome users saw a blank blue page after Google sign-in. Root cause: URLs with trailing slashes were not normalized in tool lookup, and a null tool state was rendered as an empty page instead of a loading indicator.Empty Stream Guard: Added empty-stream protection to all 6 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 operations.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 it easier to write and proofread long instance descriptions.Assistant RAG Collections: Assistants can now be scoped to specific document collections in the IP Vault. When configured, the assistant’s RAG retrieval only searches uploads from 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 has access to. Set an assistant to “Selected Tools Only” mode and pick specific tools from the institution’s enabled set. A filter toggle lets you quickly view only selected tools or expand to the full list.Assistant-Scoped RAG Search: The standalone RAG Search tool and MCP search-instance-rag endpoint now accept an optional assistant ID to scope results to that assistant’s configured collections. Search results display collection names and the number of files searched, even when no results are found.ANN Vector Search: Switched RAG vector search from Exact Nearest Neighbor (ENN) to Approximate Nearest Neighbor (ANN) with HNSW indexing for significantly faster retrieval at scale.PDF OCR Fix: Fixed an issue where image-based PDF pages failed OCR processing because the original PDF file extension was passed to the vision API instead of the converted PNG 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 are now routed 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 now default for Google real-time conversations.Admin UI Personalization Reorganization: Reorganized the Personalization tab with a dedicated Convo (Real-Time Speech-to-Speech) section, improved Document Summarization placement, and sub-section visual hierarchy with background styling.Voice Preview Links: Added direct preview links to OpenAI TTS Playground and Google AI Studio for voice audition before selection.ElevenLabs Credential Warning: Warning pill now appears in the Personalization tab when ElevenLabs Agent ID or API Key are missing, guiding admins to the Configuration tab.Lightbox Stability: New preventBackdropClose prop for modal components prevents accidental dismissal during critical workflows.Canvas Course Filtering: Fixed an issue where conversation history was not filtered by course context when launching Pria via Canvas LMS SDK integration.Canvas Section Reorder: 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, ensuring 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: Users can now choose between three RAG modes — Normal (default RAG-augmented responses), Search Only (returns only RAG results without AI generation), and Disabled (skips RAG entirely for pure LLM responses). The mode is available in the Pria sidebar, as a toolbar cycle button, and in the user profile page.Admin Histories Pagination & Filtering: The admin Histories page now uses server-side pagination with infinite scroll, replacing the previous approach of loading all records at once. Added a Course Name filter to narrow results by LMS course. 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, replacing the previous approach 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 all paginate on the server with infinite scroll (or ReactTable manual paging for Questions). A “Showing X of Y” record count appears above each table, and additional pages load automatically as you scroll down.

March 2026

Mistral AI Native Provider: Added Mistral AI as a first-class provider with native SDK integration — conversation, vision, reasoning, code, multilingual, TTS, STT, embeddings, and moderation models are now routed through the Mistral API directly instead of the OpenAI compatibility layer, improving performance and enabling full feature support
YouTube URL Import: Paste a YouTube video URL into the URL uploader and Pria automatically extracts the transcript with timestamps, video metadata, and thumbnail. Videos without captions can still be imported with Skip Indexing enabled. Re-ingesting a YouTube file refreshes the transcript and thumbnail from the source
Reprocess Content: Redesigned Reprocess Content dialog with dedicated action cards for re-embedding and re-ingesting, Skip Indexing option for re-ingest, and improved error handling that restores file status on failure
Skip Indexing: Upload files, URLs, or clipboard content to your IP Vault without generating AI embeddings — 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 via Reload or Reingest
SSO Provider Management: Admins can now configure OAuth 2.0 Single Sign-On providers directly from the Admin dashboard — define identity provider endpoints, client credentials, field mappings, and assign them to institutions. Users authenticate at /sso/{slug} and are automatically logged in and enrolled. Includes connectivity testing, error handling for disabled or missing providers, and dark mode support on the SSO login page
File Preview: Enhanced fullscreen file preview with HTML document rendering (Preview and Source views), print-to-printer support for all file types, and an open-in-new-tab button for viewing files externally
Guest Mode: Institutions can enable a minimal guest interface that hides the sidebar, file uploads, assistant switching, and profile menu — ideal for demo 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 throughout the UI now show a smooth 3D icon-flip animation instead of popup alerts, with print-optimized stylesheets for clean 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 between WebRTC and WebSocket connection methods for ElevenLabs voice conversations
RAG Search Timing: Search results now display RAG query duration for performance transparency
Google Document Upload: Paste Google Docs, Sheets, or Slides URLs directly into the file uploader — content is read via Google OAuth instead of scraping raw HTML. Supports personal and institution-level Google authorization with context-aware error messages guiding users to connect their account when needed
Google Slides Reader: Google Presentations are now fully supported in URL uploads, extracting text from all slides including shapes, tables, and nested groups
RAG Segments Panel: Enhanced segment management with content sanitization, metadata editing, and one-click reprocess workflow
Gemini Live Stability: Improved WebSocket reconnection handling for Gemini voice conversations with automatic retry on unexpected disconnects
Smarter Conversation Memory: Pria now intelligently compacts older conversation history into concise summaries while keeping your most recent exchanges verbatim, so longer conversations stay coherent and responsive without consuming excessive credits. Compaction activates automatically when Remember History is set to 6 or higher, and summaries are cached between requests for faster follow-ups
Streaming Performance: Dramatically smoother AI response streaming — delta-based text appending, frozen text caching, and TTS gating reduce per-chunk processing by ~12x, 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: Integrated 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 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_file tool, 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 loader animation and streamlined follow-up interaction UI
Multi-Provider Embeddings: RAG document embeddings now support Bedrock Titan, Google Gemini, and OpenAI models, selectable per institution
RAG Chunk Enrichment: Smarter document processing with per-chunk keyword extraction for more accurate search results
Embedding Segments: New API endpoint and 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 .mov recordings directly — Pria automatically extracts audio and converts to a universally supported format before transcription, so iPhone voice memos and QuickTime exports work seamlessly across all AI providers
API Key Authentication: Admin users 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 functionality for non-admin users to protect sensitive AI-generated content — text selection and copy shortcuts are blocked while admins retain full access
Animated Loading Screen: Redesigned loading experience with animated Pria logo, radial fade effect, and smooth fadeIn transitions
Bundle Optimization: Eliminated 1.4MB monolithic JavaScript bundle through code splitting for faster page loads
Security: Membership status enforcement blocks inactive or pending institution members from AI interactions; all active sessions revoked on password change; Facebook login migrated to server-side OAuth2; password policy enforcement; SVG sanitization; vulnerability remediation
Dark Mode: Fixed select dropdown and input background inconsistencies across dark mode
File Downloads: Corrected file extension mismatches in downloaded files
Deep Research: Reasoning effort automatically clamped for deep-research model configurations
Uploads Access Control: Three-tier security model for file access — owner, institution members, and shared access levels
iOS Safari: Fixed keyboard toolbar overlapping chat input area on iOS Safari

February 2026

Gemini Flash Live: Added Google Gemini 2.5 Flash as a third real-time voice provider in Convo Mode via WebRTC, 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 as 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 model and set as default for Convo Mode
Direct AI Providers: Introduced native Anthropic and Google GenAI providers for direct model access, replacing the OpenAI compatibility bridge for improved performance and 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, and deprecated model indicators in the model catalog
File Collections: Organize files into nested collections with drag-and-drop, breadcrumb navigation, and animated transitions. Bulk edit mode for multi-select move, delete, and reprocess operations
Files Redesign: Rebuilt the Files interface with a virtualized grid, paginated infinite scroll, live upload progress indicators, and enhanced file cards with loading states and 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 split into time-based chunks and transcribed in parallel with speaker diarization support, raising the upload limit to 150MB
Excel Legacy Support: Added support for older .xls Excel spreadsheet format
History Compaction: LLM-based 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, visible 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 with garbage detection heuristics for code-heavy or low-value embeddings
RAG Error Recovery: Files stuck in processing are now automatically marked as failed with a red indicator. Stale uploads (>15 minutes) offer one-click Delete or Reprocess actions
Edit Assistant Redesign: Redesigned the Edit Assistant lightbox with improved variable management UI 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) 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 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 user onboarding with auto-loading of eligible users
Platform Monitoring: Enhanced platform-level session monitoring with SSE tracking, per-transport hourly stats, and persistent client session history
Admin Filtering: Added AccountSelect filter across all admin list pages for easier multi-account management
Security Hardening: Server-signed HMAC-SHA256 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 support 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 requiring a manual save action
Feedback Redesign: Redesigned 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 enrollment page with keyword search capability
Performance: Optimized file queries, assistant loading, SSE connection handling, and session chart aggregations for faster page loads and reduced admin timeouts

January 2026

Dark Mode (Beta): Release of the dark mode support in 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 improved visual hierarchy and screen real estate optimization
Files and Assistants Layout: Repositioned Files and Assistants under the text entry bar for a more compact, streamlined display. Redesigned file and assistant cards
Profile Picture Upload: Added ability to upload profile pictures from webcam or file, with automatic RAG indexing
Sidebar Enhancements: Collapsible sidebar tray, direct tool enable/disable toggle, and 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 Q&A streaming endpoint for real-time response delivery
Prompt Optimization for 2026 Models: Periodic optimization of system instructions to align with evolving model capabilities (GPT-5.2, Claude 4.5, Gemini 3.0 Flash). Prioritized personality-first responses to create more immersive Digital Twin experiences, balancing conversational tone with technical precision
Tool Guidance Toggle: Added a new switch to reduce instruction verbosity for general conversations while maintaining comprehensive guidance for advanced research queries. Optimizes token usage and response latency without sacrificing quality
Model-Aware Response Tuning: Enhanced instruction stack to leverage native capabilities of December 2025+ models, including improved reasoning, reduced hallucination rates, and better context retention across long conversations
Voice Selection: Enhanced voice picker for Convo mode with improved selection experience
SDK Hardening: Improved SDK reliability with inactive instance handling and membership validation
File Cleanup: Added automatic cleanup routine to remove orphaned files from deleted users and institutions

December 2025

Open AI GPT 5.2: Added support for latest GPT 5.2 model from Open AI
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 for maximum reuse and consistency.
Account-Level RAG Files: Documents can now be promoted from Instance to Account level, enabling every digital twin in the account to access a shared knowledge base.
History Graph: Now defaults to the last 12 weeks with daily data rendered 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 embeddings, content summarization, and image analysis during file/URL uploads.

November 2025

Sonnet 4.5: Added support for Claude Sonnet 4.5 model and set as platform default
All Anthropic Models: Upgraded plan to support large input context up to 1 million tokens
Tool Upgrade: Improved support for call_canvas and search_canvas to enable simultaneous execution
Content Length Returned by Tools: Tools now return up to 60,000 characters (increased from 20,000), enabled by expanded input context size
Assistants Features: New options to disable RAG searches when executing assistants and run assistant instructions without system context (Platform and Digital Twin instructions)
Pria JS SDK: Added pria.isReady() function that returns true/false to indicate whether Pria is ready to receive messages
File Import Enhancements: Improved support for ZIP and IMSCC file imports

October 2025

Shared Memory: Added support for shared memory functionality. Shared memory allows storing information that is accessible across all users of the same digital twin, enabling cross-user data persistence and collaboration.
Memory Tools Enhancement: Renamed set_in_user_memory to set_in_memory and get_from_user_memory to get_from_memory to support both user-specific and shared memory operations with unified tool interfaces.
Privacy Policy Enforcement: Introduced account-level privacy policy settings that require users to accept terms before accessing Digital Twins, ensuring compliance and user consent.
Selective File Upload Authorization: When file upload functionality is disabled for users, administrators can now specify email domain suffixes to authorize selected user groups for file upload capabilities.
Selective Assistant Authorization: When Assistant functionality is disabled for users, administrators can now specify email domain suffixes to authorize selected user groups for Assistant access.
Favorite Digital Twins Priority: Digital twins marked as favorites now appear first in the gallery carousel, providing easier discovery and quick access to frequently used digital twins.
OCR for Image-Based PDFs: Image-based PDF files (containing no extractable text content) now automatically undergo Optical Character Recognition (OCR) processing for text extraction and indexing.
Enhanced Date Range Selection: Administrators can now filter history and session records using a comprehensive date range picker, improving record management and data analysis capabilities.
Documentation Updates: Comprehensive updates to documentation sections covering RAG (Retrieval-Augmented Generation), Bring Your Own Token (BYOT), call_canvas and search_canvas tools, and additional features available at https://docs.praxis-ai.com

September 2025

Anthropic Sonnet 4.5: Support for Claude Sonnet 4.5 (20250929) with 200K context window, 64K output tokens, and enhanced multimodal capabilities including vision, document analysis, and tool use.
MCP Server: Introduced MCP Server in Praxis-AI Middleware to let any AI solution call your Digital Twin.
Security & Performance: Improved RAG security and performance.
REST APIs: Expanded functions in 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 forced to reset their password on next login.
Logged Remote Calls: Remote connector MCP calls now logged in Tools Details.
Gepetto Digital Twin: Improved AI Portrait Artist to generate Digital Twin avatars.
Moderation by Default: Moderation service is now enabled by default when creating new Digital Twins

August 2025

Connector MCP: Added comprehensive support for Remote Model Context Protocol (MCP) servers, individually configurable for each Digital Twin instance to enable distributed AI capabilities
Custom LLM Integration: Implemented support for custom Large Language Models (LLMs) with per-Digital Twin configuration, allowing organizations to deploy specialized AI models tailored to specific use cases (ex: Vertex, etc.)
Digital Twin Gallery Enhancements: Significantly improved user experience with intuitive navigation arrows and resolved display flickering issues on macOS systems for smoother interactions
OAuth Verification for Canvas Integration: Enhanced security architecture for Canvas Theme SDK access by implementing mandatory OAuth 2.0 authentication workflow, ensuring secure user verification before Digital Twin initialization
Streamlined Digital Twin Creation: Empowered administrators with a simplified 2-step Digital Twin creation process directly through the Gallery interface, reducing deployment complexity and time-to-value
Support for Mermaid diagrams and charts: Added support for Mermaid Diagramming and Charting

July 2025

User Experience: Enhanced digital twin loading for standard users
Branding: Implemented branding feature for personalized login/signup
Text Customization: Externalized activation and joining messages
Logo Branding: Added logo parameter for login sequence customization
Animation: Added animated picture support for Digital Twins

June 2025

Token Management: Improved Canvas faculty token refresh system
Digital Twins: Added gallery for twin selection and management
REST Tools: Custom REST-enabled tool creation for instances
Email Notifications: Enhanced approval workflow notifications
Speech Features: Removed deprecated react-speech component
Accessibility: Improved contrast ratios and ARIA labels
Substitution Variables: Developed prompt variable management system
Email Tool: New email functionality implementation
Conversation Persistence: Added save/restore last conversation feature

May 2025

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

April 2025

Credit Transfers: Enhanced account-to-instance credit management
File Tools: Added get_user_files tool for document collection access
Instance Management: Developed comprehensive admin management system
Performance: Improved file loading with asynchronous summaries
Pricing: Decreased costs across personal accounts

March 2025

Email Processing: Implemented postal-mime package for content parsing
Node.js Upgrade: Updated to Node 17 for AWS SDK compatibility
Streaming: Enhanced conversation streaming capabilities
Canvas Scoping: Added course ID filtering for history records
Assistant Management: Made user assistants admin-visible for sharing

February 2025

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 SASS NPM module issues and file system errors

January 2025

Accessibility: CSS improvements for epileptic users
Canvas SDK: Updated to version 1.3 with enhanced user ID handling
Security: Updated VisionLoft SDK to address vulnerabilities
Navigation: Improved UI layout for large screens
API Integration: Enhanced Canvas API access with call_canvas tool
SDK Features: Added browser messaging capabilities and documentation

February 2024

Scientific Literature: Added CrossRef tool for academic research support
Real-time Communication: Implemented OpenAI WebRTC APIs with UI
Voice Features: Enhanced voice selection capabilities

January 2024

Canvas Integration: Captured global ID for Canvas oAuth tokens
User ID Updates: Relaxed user ID requirements to support string format

December 2024

Data Compliance: Implemented DPA compliance features for data deletion and recovery
Canvas Integration: Added discussion bot plugin
Token Management: Enhanced Canvas token refresh capabilities

November 2024

Database Improvements: Enhanced DocumentDB aggregate lookup support
Retry Logic: Implemented Bedrock retry mechanisms
Analytics: Added tool usage tracking and metrics
Instance Management: Added ability to disable specific assistants per instance

October 2024

Email Notifications: Added low credit alerts for administrators
Location Services: Enhanced GSearch with location-based results
AWS Integration: Switched to AWS as default platform for document summarization
Course Management: Added courses in sidebar for institutional scoping
Canvas tool: Enhanced LLM interpretation and display capabilities
Student Analytics: Developed feature to analyze student questions
Assistant Features: Launched new Assistant functionality and Course Syllabus Generator
Bug Fixes: Resolved input size limitations and tool usage issues
Admin Enhancements: Multi-select instance support and Gemini Flash 1.5 implementation

September 2024

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 mini models
Scalability: Fixed entitlements listing performance issues

May 2023

UI Improvements: Various interface enhancements and bug fixes

April 2023

UI Enhancements: Implemented anchor text conversion for links
Copy Functionality: Added copy input to clipboard feature
Timeout Resolution: Fixed response timeout issues (25-30% timeout rate resolved)
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: Implemented HTML image tag support

Product Updates

Documentation Index