What's New Ask Shiv about a Manas's documents Open a chat scoped to a Manas and the PDFs and Word files attached to its notes are now searched and cited by page or heading. Documents from other Manases stay out of the answer. Embedding is about 2 seconds, not 12 Most text now uses a smaller, faster embedding model. Only long notes switch to the bigger one. Both give the same result, so nothing you've already saved needs to be re-embedded. Fixes Shiv gave the same answer to every question with the small Qwen3 0.6B model — fixed. Deleted notes lingered in Shiv's memory. Deleting a note now removes its search vector, summary, and graph links too. A Shiv reply that ended with no text showed a blank bubble — it now says "No response from the model." Queued knowledge extraction could stall for minutes on an idle phone — fixed. Removing an unreachable relay in Settings could hang — fixed. A failed switch between the two embedding models could silently drop notes from the queue — they are now retried.

The biggest Shiv release yet. Shiv can now understand your documents and photos, translate notes entirely on-device, and explore your knowledge through a completely redesigned graph — all while keeping your data on your device. * Ask questions about documents - Attach PDF or Word files to notes and ask Shiv questions about them. - Answers include sources, with PDF page numbers or Word headings. - Tap a source to jump directly to the relevant passage, highlighted for you. - Scanned documents, photographed pages, and images are processed entirely on your phone. - English and Hindi text in scans, photos, and documents are supported. - Photos without text can also be searched by what they contain — like “a dog on a beach.” * Translate any note - Translate notes into another language using on-device AI. - No internet connection or cloud processing required. * Redesigned knowledge graph - A clearer, Obsidian-inspired graph with improved layouts. - Connected nodes light up around the selected note. - Drag nodes and their neighbours follow naturally. - Large graphs open automatically laid out and fitted to the screen. - Search your graph and navigate through a note's connections one by one. * Faster AI on Android - Shiv now uses the phone's GPU on supported Android devices. - First responses appear about 2.5× faster. - Model downloads can continue in the background with notifications. - Devices that cannot run the model on the GPU now automatically fall back to the CPU. * Better Hindi & Hinglish search - Numbers, names, and codes now contribute alongside semantic meaning. - Hindi spelling and digit variants are handled more consistently. * Fixes & improvements - Graph now refreshes correctly after replying from a note panel. - Reloading the graph while viewing a Manas no longer unexpectedly exits it. - Fixed graph nodes that could not be tapped. - Fixed Shiv occasionally repeating its prompt at the beginning of an answer. - Improved document search so all stored passages can be found reliably.

What’s New This update brings a more reliable and seamless AI experience to UNIUN. New Features - QR Sign-In: Sign in to UNIUN Web by scanning a QR code from your phone—no password or key required. - Ganas Cloud Models: Pin a UNIUN Cloud model specifically to Ganas, independently of the model used elsewhere in the app. - Automatic Model Cleanup: Detect and remove incomplete or leftover model files to reclaim previously unused storage. Improvements - Unified Model Selection: Shiv, Nataraj, and Ganas now use one consistent model picker. - Onboarding: Interest suggestions are now loaded from the UNIUN backend, with loading and retry states. - On-Device AI: Updated the AI engine for improved reliability during extended use. - Model Switching: Changing models now takes effect immediately and safely, even during active generation. - Ganas: Agents without a pinned model now automatically follow the app's currently active AI backend. Bug Fixes - Fixed model pickers displaying internal routing IDs instead of model names. - Fixed Nataraj showing an incorrect device compatibility error when no idea could be generated. - Fixed cloud AI generation occasionally hanging or failing when requests overlapped. - Fixed errors when switching or deleting models during active AI generation. - Fixed Shiv responses becoming blank when leaving a chat while the AI was still responding. - Fixed empty Shiv conversations appearing in chat history. - Fixed duplicate connections occasionally being counted in Nataraj's knowledge-graph search.

This release expands UNIUN's inference, identity, and offline networking capabilities. ### Added - **UNIUN cloud inference.** Use hosted AI models from Shiv with built-in model selection, credit usage, and mobile payment support alongside on-device inference. - **QR identity recovery.** Sign in on another device by scanning a secure QR code instead of manually transferring identity credentials. - **Offline mesh sync.** Nearby devices can discover each other over local transports and securely synchronize notes, follows, groups, Manas collections, Ganas, and other user-owned data. - **Surrounding notes.** Discover notes shared by nearby mesh peers in a dedicated feed. ### Changed - Onboarding interest recommendations are now loaded from UNIUN's service so suggested accounts can be updated without an app release. - Expanded repository, gateway, mesh, and integration test coverage. ### Fixed - Improved cloud-model selection, credit accounting, inference UI, login, and sync reliability.