--- title: "AssistantGeneral Changelog — Beta Notes" description: "Every AssistantGeneral beta build and what changed in it: recordings searchable by the moment, profession workspaces, practice packs, helpers and more." url: "https://royalsoftworks.com/products/assistant-general/changelog/" source: "https://royalsoftworks.com" format: "markdown" note: "Markdown rendering of the HTML page at `url`. Same content, same canonical URL." --- Changelog # What changed, build by build. AssistantGeneral is in a private beta. Every build is listed here with what it added, what it fixed, and what it still does not do — the last part included on purpose. Private beta. Builds go to testers directly; ask to join. [Join the beta](https://royalsoftworks.com/contact/?product=assistant-general&tier=beta) Newest build AssistantGeneral 2.0.0-beta.4 12 October 2026 Search your recordings by the moment, answers from your own documents first, and faster writing. Private beta, Windows only No telemetry in any build [Ask to join the beta](https://royalsoftworks.com/contact/?product=assistant-general&tier=beta) Read what's in it ## 2.0.0-beta.4 Next beta 12 October 2026 Search your recordings by the moment, answers from your own documents first, and faster writing 2 New 5 Changed 4 Fixed Recordings and videos can now be searched like documents: an answer cites the minute and second, and clicking it plays the recording from there. Questions your documents can answer are answered from them first, in one step instead of several, and every source in an answer is a visible link. The built-in model writes faster on the same computer, a slow computer is no longer asked to do work it cannot afford, and a long list of rough edges found in a whole-app test is gone. Worth knowing about this build - Still a private beta, Windows only: builds go to testers directly and are not on the download page. - The faster writing was measured on one test computer with an older NVIDIA graphics card. It has not yet been measured on Intel or AMD built-in graphics, and it applies to the built-in model only. - Finding a moment by what is shown or what it sounds like needs extra files (about 500 MB), downloaded from Settings. Without them, recordings are searched by what is said in them. New ### Recordings and videos you can search by the moment Audio and video files in a knowledge base are cut into 30-second moments. Ask about something said in a meeting recording and the answer cites the minute and second; click the time and the recording plays from that point, inside the app. With the optional picture-and-sound files, a moment can also be found by what is on screen or what it sounds like, even when nobody says a word about it. New ### Sources you can see are links Every source an answer cites carries a small icon — a document, a recording, a video or a web page — so it is obvious it can be clicked. A document opens in its own program, a recording plays from the cited moment, a web page opens in your browser. A program file is never started from a citation. Changed ### One search model for every language The model that makes documents searchable is now multilingual for everyone, so Serbian, German or Arabic documents are searched directly instead of being translated first. A knowledge base built with the previous model is re-indexed with the new one from the text it already holds — the original files are not needed. Changed ### Your documents first, in one step The assistant could reach for a web search for something your own documents already answered, and a question about your documents went through several rounds of internal planning and checking before the answer started — minutes, on an ordinary office graphics card. When your documents contain the answer, it now comes from them, with the source cited, in a single step; the web is used only for what your documents do not say. Changed ### Answers are written faster A small helper model, published beside the built-in model and fetched by the app (60 MB), guesses the next few words and the main model checks each one, so the answer is exactly the one it would have written anyway — it just arrives sooner. On our test computer that was about 1.6 times faster on its graphics card, about 1.3 times in the mode most built-in graphics use, and a little on the processor alone. It can be switched off in Settings. Changed ### A slow computer is not asked to do the slow extra The speed check at first start now changes something: on a computer it finds slow for the local model, answers skip the second fact-check against your documents, which on such a machine can add a minute after the answer is already on screen. Settings says when this is happening and can turn the check back on. Changed ### Newer engine underneath The built-in model now runs on the latest stable release of its engine (llama.cpp 0.6.0), and the optional alternative engine moves to KoboldCpp 1.122.1. Fixed ### Chats keep their settings With encryption on — the default — the second message in a chat could fail with an error, and saving a conversation dropped the chat's own settings: which documents it may read, its workspace and its practice pack. Both are fixed, and the chat file stays encrypted. Fixed ### Chats have names, and any name works A chat started by simply typing used to appear in the list as a long code. It is now named after its first question. A name with characters Windows does not allow in file names, such as a colon or a question mark, no longer breaks the chat. Fixed ### The model starts reliably, and says why when it cannot Starting the model could judge a good start a failure and fall back to a smaller setup. A model left running by an earlier session that ended badly is now stopped first; if another program is already using the address the model needs, the app names it instead of fighting it; and when the model really cannot start, the app says why and what usually helps. Fixed ### The quick window tells the truth The small always-on-top window showed the model as off even when it was running, and had no way to stop a long answer. It now shows the real state and has a Stop button, and clearing it mid-answer stops the answer instead of leaving the window locked. ## 2.0.0-beta.3 Current beta 5 October 2026 Workspaces for your profession, quick answers, and governance you can prove 10 New 4 Security 2 Changed 1 Fixed The biggest change since the beta began. A chat can now belong to a profession — law office, judiciary, hospital, clinic, HR or recruitment — and opens in a workspace built for that job, with its own everyday actions and its own finished document. On top of that sits a practice pack: the country the work is done in, with the laws normally cited there, the usual procedures and their deadlines, and the forms people actually file. Simple questions now get a quick answer instead of the whole machinery, a workflow can be described in plain words and built for you, and for organisations every outbound connection now passes one gate that an Offline mode can close. Worth knowing about this build - Still a private beta, Windows only: builds go to testers directly and are not on the download page. - Practice packs ship for four places: International (no national law named), Serbia, Germany and the United States. The US packs deliberately stop at federal law and say so — they will not guess which state you are in. - The deadline calculator skips weekends but does not know public holidays, and tells you that on screen every time. - The citation check is switched off by default. When on, it can prove a US case exists; it cannot tell you whether that case is still good law. New ### Quick answers for simple questions Every message used to pay for the whole pipeline — planning, document search, a long tool list, the answer checks — before the first word appeared; on an ordinary office graphics card a plain factual question took close to two minutes. A simple message now gets one short model call. Anything that needs your documents, a file, the internet or a tool goes to the full pipeline as before, and the chat shows it working rather than looking stuck. New ### Describe a workflow and the app builds it In the workflows dialog, describe the job in plain words — what starts it, what it reads, what it should produce — and the assistant assembles the workflow from the available steps, checks that it is valid, and repairs what does not fit before showing it to you to edit or run. New ### Workspaces, one per profession, side by side in tabs A row of tabs under the header, one per workspace, each with its own list of chats — so a lawyer can keep a matter open in a Law office tab while an HR colleague works in the HR tab on the same machine. Six ship: Law office and Judiciary (an advocate argues for a client; a judge decides between parties — same documents, opposite duty), Hospital and Clinic (a ward versus a single practice), HR and Recruitment (the people you have versus the people you are hiring). The ordinary General assistant is still there, exactly as it was, for everything else. New ### Each workspace does a job, not just a relabel Every workspace names the thing you are working on (a matter, a case, a patient, a candidate), limits the chat to the document collections that belong to it, puts the everyday actions for that job one click away, shows the evidence an answer used down the side, and produces a document you can save out of the app. The limit on which documents a chat may read is real: it is enforced where the search happens, not just hidden in the interface. New ### Practice packs: the country you work in A workspace says what job you do; a practice pack says where. A Serbian advocate files a tužba under the ZPP and has eight days to appeal a rešenje; a German one files a Klage under the ZPO. A pack carries that difference as reference data — the authorities normally cited and how they are written, the usual procedures with their stages and deadlines, the documents people produce with the sections a valid one has, and the checklists a careful professional works through. Twenty-four packs ship (six professions × four jurisdictions), and you can copy one and edit it for your own practice. New ### Six practice tools in every pack Compare two documents aspect by aspect, with which side is worse and why. Extract the fields of a known document type, each with the passage it came from. Ask a question and get an answer where every point carries the authority, the article and how it applies to your facts. See a procedure as stages with the real dates worked out. Draft a document from the facts it needs. Run a checklist against the matter, item by item. All six go through the ordinary chat, so they follow the same document limits and leave the same record. New ### A deadline calculator that never asks the AI Enter the date that starts the clock and it works out every deadline in the procedure with ordinary arithmetic on your machine — no model involved, because a confident, plausible, wrong date is the worst thing a legal tool can produce. Every stage is counted from the date that actually triggers it, which is where hand-chained calculations usually go wrong. New ### Citation check for legal answers (off by default) A new final check looks at the cases and statutes an answer cites and asks a public legal database whether they exist. A citation that the source affirmatively reports as not found sends the answer back to be corrected. A source that is offline, slow or has no coverage never condemns a citation — silence is not evidence. The first source covers US case law; it proves a case is real, not that it is still good law. New ### A guided first run, and a tour that points at the real screen First launch is now one short sequence — welcome, licence, language, what you work on, engine, models, done — instead of a chain of separate pop-ups. The old step-by-step tutorial is replaced by a tour that highlights the actual buttons and opens the real dialogs in a safe demo form, so you learn where things are rather than reading about them. The whole setup can be run again from Settings → Help. New ### Administrators decide which workspaces and packs each team gets For organisations, workspaces and practice packs are part of the policy an administrator sets, alongside new controls for helpers, the citation check, local engines, model downloads, backups and app updates. Policies written before this release keep working exactly as they did. Security ### One gate under every outbound connection, and an Offline mode Every connection the app opens to the outside world — model providers, connectors, downloads, the licence and update checks — now passes through a single gate. Offline mode closes it to everything beyond your own network; an administrator can instead allow a short list of hosts. A blocked name is refused before it is even looked up, and every host that was reached is counted in a daily record on the machine. Security ### Employees see what their administrator did When an administrator runs a tool on an employee's machine, pushes or pulls a file, or changes their settings, the employee gets a notification that survives restarts, under a distinct Administrator label. The remote database query, which let an administrator read an employee's indexed documents without telling them, has been removed entirely. Security ### A company policy stays in force, connected or not An employee's policy is now stored on their machine and applied at every start, whether or not the company's admin server is reachable, so a managed desktop is never left without one. Organisations also get a read-only auditor role, per-connector and per-tool action limits, and an activity record that is always on for enterprise installs. Security ### A licence key works in AssistantGeneral only Activation now names the product it is for, so a key bought for another Royal SoftWorks app is refused instead of unlocking this one. Changed ### A calmer, denser window The header, footer and chat header are all shorter, the engine on/off switch moved to the main header where it is always visible, menu names were cut to one word each, and Settings → Core is split into Model, Answers, Knowledge and Integrations tabs instead of one long page. Changed ### The answer self-review stays off The pass in which the assistant rereads its own draft before showing it was switched off in beta 2 while a topic-mixing bug was fixed. The bug is now fixed in the review itself, but on a local model the extra pass costs more time than it saves, so it stays off in this build. Answers that use your documents or tools still have their facts checked against what was found. Fixed ### Disabled buttons that were not actually disabled Across the app, a button drawn as greyed-out could still be pressed, including some that a policy was meant to lock. Every one now genuinely refuses, and a build check stops the mistake coming back. Where a button is blocked because a field is empty, pressing it now tells you which field instead of doing nothing. ## 2.0.0-beta.2 Superseded 20 September 2026 Helpers, backups, and a first run that works anywhere 8 New 2 Security 1 Changed 2 Fixed A month of work between two tester builds. The assistant can now hand parts of a big request to helpers that work separately and report back. Your data is backed up on a schedule, encrypted with a password only you hold. Setup downloads what your particular computer needs and has a way through when the internet is blocked. And a long list of things that looked finished but were not wired up now are. Worth knowing about this build - Windows only. No Linux build yet. - Not code-signed, so Windows shows a SmartScreen warning. Automatic updates are not switched on; new builds arrive as a fresh download. - Helpers (below) are on by default and are the least-proven part of the build. - WhatsApp Personal is listed but cannot connect; WhatsApp Business works. Company single sign-on is switched off for the beta. New ### Helpers: big requests split into separate pieces of work When a request has independent parts — research three suppliers, compare two contracts, summarise a folder — the assistant can hand each part to a helper that starts with a clean slate, works on its own, and returns a short report. The benefit on an ordinary computer is focus rather than speed: forty pages of web results stay in the helper's memory instead of crowding the main conversation. It checks how many jobs your engine can genuinely run at once and queues the rest, always keeping one free for the conversation you are in. You can set up your own helper roles. New ### Scheduled, encrypted backups Chats, knowledge bases and settings can be backed up on a timetable, with old copies pruned automatically. Each archive is encrypted with a password you choose — deliberately not with a key stored on the computer, because a backup that only opens on the machine that died is not a backup. A backup missed because the computer was off runs shortly after the next launch, and asking for encryption without a password is refused rather than quietly writing an unprotected copy. New ### Live CPU, memory and graphics-card monitor A small gauge shows how hard the machine is working while the assistant thinks, so a slow answer on a busy computer looks like what it is rather than like a hang. New ### Reads email files, audio recordings and saved web pages Saved emails, audio files (transcribed on your machine) and local HTML pages can now go into a knowledge base alongside PDFs and Office documents. New ### Search presets for legal, technical and general documents The document search has around twenty settings. A preset moves them together for the kind of material you have — a statute book wants different treatment from an API manual — with the reasoning shown next to it. New ### A choice of ranking model, from 22 MB to 545 MB The model that puts the best search results first is now a choice in Settings: a large multilingual one, a mid-sized English-and-Chinese one, or a very small English one for modest machines. It downloads in the app and applies immediately, with nothing to re-index. New ### The engine installs itself on any computer — or from a file you bring On first run the app picks the engine build that matches your graphics card and installs it. Where a corporate proxy or an offline network blocks that, it now says so and offers the manual route: download the engine yourself and point the app at the file or folder. The models themselves now come from a public source, so a fresh install no longer depends on our servers at all. New ### Three features that were described but unreachable Editing a trigger after you create it (before, a typo in a schedule meant deleting and starting again), exporting and importing whole bundles of workflows, and scheduling Instagram posts all had the machinery and no button. They have buttons now. Security ### The PIN lock is a real lock The launch PIN used to be enforced by the interface. It is now enforced by the part of the app that holds your data, so nothing — including your settings — is handed over until the PIN is entered. Security ### Company single sign-on is switched off for the beta The sign-on and user-provisioning services for organisations work, but there is no screen yet for an IT team to configure them. Rather than ship something configured by nobody, they are off in this build and turned on per organisation during a pilot. SAML sign-on is not offered at all until its signature checking is complete. Changed ### A third less preparation before every answer The instructions the assistant reads before your question went from about 12,900 tokens to about 7,900 without losing any tools. On a local graphics card that is time you get back on every single answer. Fixed ### Changing the subject no longer drags the old one along Ask about the weather, then about something else, and earlier builds could keep answering the weather too. The assistant now notices when you have started a new task, sets the old plan aside, and stops steering you back to it. Fixed ### Settings → Core is fully translated One hundred and fifty-eight pieces of text on the main settings page were showing in English in every language. They are translated now, and a build check catches text that is translated but never displayed that way. ## 2.0.0-beta.1 Superseded 20 August 2026 The first tester build 7 New 2 Security 1 Fixed The first build handed to outside testers. Most of the work was making the product fit to be installed by someone who is not us: a first run that starts with your language, activation that actually checks a licence, a way to receive updates, and an app that stays up when something goes wrong. It also brought right-to-left and Arabic support up to the standard of the other nine languages. New ### First run starts with your language Before anything is downloaded you choose the interface language, and that choice picks the models: a multilingual search model for non-English documents, and an Arabic-optimised chat model when you choose Arabic. New ### Full right-to-left support, and better Arabic documents Arabic now lays out right to left throughout, mixed Arabic-and-English text reads in the right order, and text pulled from Arabic PDFs comes out in the right sequence. PowerPoint files can be read into a knowledge base. New ### Scanned pages read by the model you already have Text in scans and photos is read by the vision-capable model that answers your questions, when one is loaded — no separate recognition program to install, and it improves whenever the model does. New ### Slide decks with a preview before anything is saved When the assistant builds a PowerPoint deck, each slide appears in a preview and you approve or reject slides one at a time before the file is written. New ### Switch the search model without breaking your knowledge base The model that turns documents into searchable form can be changed — for example to a multilingual one — and the knowledge base is rebuilt to match rather than silently mixing two incompatible indexes. New ### A gauge for how much the conversation can still hold Every model can only keep so much of a conversation in mind. A gauge now shows how full it is, older turns are summarised at a point you choose rather than cut off at the limit, and the chat tells you when that happens. New ### Update checks and a licence screen The app can find and offer its own updates, shows a licence agreement on first run, and checks a licence key when you enter one. A key that has lapsed drops you to the free tier without deleting it, so renewing restores everything; an offline machine is never locked out for being offline. Security ### Web pages the assistant reads are kept at arm's length Pages fetched from the internet are now rendered in a separate, locked-down window with no access to the app, your files or your settings. Security ### A tamper-evident activity record Each entry in the audit log is chained to the one before it, so an edited or deleted entry shows up as a break. Retention periods and several missing event types were added. Fixed ### The app stays up when something goes wrong An unexpected error used to close the app. It now writes a short entry to a local crash file — a timestamp and the technical detail of the error — tells you, and keeps running. During the beta ## How a build reaches a tester. 01 ### Sent to you directly Beta builds are not on the download page. Testers get the installer and a full licence key from us, and each new build arrives the same way until automatic updates are switched on. 02 ### About 6 GB on first run The installer holds the app. The AI engine and the model are downloaded on first launch, matched to your graphics card — or installed from files you bring, on a network that blocks downloads. 03 ### Nothing reported back There is no telemetry and no usage log. If something breaks, the app writes a short local crash file with no chat content in it, and a tester decides whether to send it to us. ## Want to test it on your own work? We are adding testers who work with documents that cannot leave the building — law, medicine, HR, finance. Tell us what you do and we will send a build. [Ask to join the beta](https://royalsoftworks.com/contact/?product=assistant-general&tier=beta) [See what it does](https://royalsoftworks.com/products/assistant-general/features/)