Repository navigation
v0.1.63
Which file do I need?
The downloads are at the very bottom of this page, in the Assets list underneath all these notes. Scroll all the way down, and click Assets if the list is folded. You need exactly one file:
| Your computer | Download this ONE file |
|---|---|
| Windows 10/11 | HilbertRaum-0.1.63-portable.exe |
| macOS (Apple Silicon) | HilbertRaum-0.1.63-mac-arm64.app.zip |
| Linux | HilbertRaum-0.1.63.AppImage |
The other files are NOT the app: the llama-runtime-*.zip files carry the AI engine for offline/air-gapped installs only (normally the app offers that same download built in), SHA256SUMS.txt is for optional download verification, and the two Source code archives are the source for developers, not a runnable app.
This download is the app only — a working chat needs two more downloads, both offered on the AI Model screen inside the app: the AI engine (llama.cpp — the screen shows an install banner until it is present; without it, models run in demo mode with simulated answers) and an AI model of your choice. Two optional downloads are offered in the app only when you need them: the knowledge-pack tools, and the text-recognition (OCR) files for scanned PDFs and photos (German + English, about 4 MB; offered on the scan or photo in Documents and on the AI Model screen). Every download asks first and is SHA-256-verified. Offline installs: the llama-runtime-*.zip assets below hold the same verified engine build — unzip into the runtime/llama.cpp/ folder on the drive. Repo users can provision a drive up front instead (README step 2). Unsigned build: see the SmartScreen/Gatekeeper notes in docs/troubleshooting.md.
Added
- A question that got no answer can be sent again with one click. When an answer fails because
the AI model crashed, ran into an error, or was stopped before it wrote a word, the question now
shows Send again underneath. It asks the same question again, in chats and in document chats,
without adding it to the conversation a second time. Before, you had to type or paste the question
again (#613). - Scanned pages inside an ordinary PDF can now be made searchable. A typed letter with a
scanned signature page, or a report with a scanned attachment, used to index its typed pages and
leave the scanned ones silently out of search. The document's row now says how many pages are
scanned without readable text yet, and Make searchable (OCR) in its ⋯ menu reads just
those pages; the typed pages stay as they are, and answers cite the scanned pages by page. A PDF
added before this update shows its scanned pages after one Re-index (#575). - Text recognition tells you when it was unsure. When a scan or photo reads poorly (blurry,
skewed or handwritten, or in a language other than German or English), the document's row says
so quietly, and its preview names the likely causes next to a Read again (OCR) button (#538). - A damaged AI engine can be installed again from the AI Model screen. When some of the
engine's own files on the drive are missing or damaged, the banner now offers Install the AI
engine again instead of a manual "delete the folder and start over". It downloads a fresh,
checksum-verified copy, replaces the broken one and starts your model again; the voice engine's
note offers the same. Like every download, it needs Allow internet access… and a drive that
permits downloads. If even the fresh copy will not start, the banner says its files are probably
not the cause (on Windows, usually an outdated Visual C++ Redistributable) (#532). - The app tells you when the drive's AI engine is out of date, and updates it. A drive set up
before an app update keeps its older AI engine (and voice engine). The AI Model screen now
shows a quiet "An update for the AI engine is available" with an Update button, and
Settings → Diagnostics lists the versions. The new version downloads and is checksum-verified
while your model keeps answering; then the models pause for a few seconds while the files are
replaced, and your model starts again. The CPU fallback build is updated in the same step. A newer
engine than the app expects is never replaced with an older one (#516). - Installing the AI engine no longer asks you to stop a model in demo mode first. A model
answering with simulated replies does not use the engine, so the install goes ahead, and the
model then starts on the real engine by itself (#532).
Changed
-
Questions about a short document now read the whole document. When a chat answers from one
document that is short enough for the AI model to read at once (about 3 pages with the smallest
models, about 7 with most, more if you raise the context size), the app now gives the model the
whole document, in order, instead of the few passages that best match the question. Answers no
longer miss part of a short letter, and the line under the answer says Covers the whole
document. A chat with several documents, or with knowledge packs, works as before (#643). -
Some display preferences are reset once after this update. Because the app's screens moved
to their own address, the browser engine treats them as new and forgets four display
preferences once: the unlock screen appears in your computer's language until you unlock (it
then follows your language setting again), and the chat list and the Documents side panels open
in their default state. Your workspace, documents and settings are not affected (#560). -
Help for Linux when the AppImage does not start. The user guide and the README now give the
exact commands to make a downloaded AppImage executable and run it, and Troubleshooting has a
new entry, "Linux: the AppImage does not start". It covers the FUSE 2 library that newer
distributions leave out (which package to install on Ubuntu, Debian, Linux Mint and Fedora, and
how to start the app without it when you have no administrator rights), and why a prepared
drive is started with its launcher rather than the AppImage itself. -
The Linux and Windows requirements are written down. Linux needs Ubuntu 22.04, Debian 12 or
Fedora 36, or newer; Windows needs Microsoft's Visual C++ runtime, which most PCs already have.
The README, the user guide, the drive's "READ ME FIRST" and a new Troubleshooting entry, "The AI
engine can't run on this computer", give the exact package to install (#530). -
The formatting of answers is updated. Streamdown, which lays out the assistant's answers,
moves to version 2.6. It no longer brings along the Mermaid diagram library, which HilbertRaum
never used or shipped, so about 110 unused packages leave the project. While an answer is still
being written, a few edge cases format better: a product like2*3*4no longer shows a stray
asterisk after it. Answers otherwise look as before (#550). -
The AI Model screen says where each automatic model is used. The models that work on their
own (document search, voice, translation and the vision model) have no Use button, and the
one sentence that explained this was hidden under Technical details. It now shows on each
model's card. The vision model's card says it is used on the Images screen and, once
downloaded, has a Go to Images button (#539).
Fixed
-
A "how much" question gets an answer, and a sum over a short document gets a total. Once a
document had a deep index, "How much is the monthly fee?" was answered at once with every amount in
the document. "How much did I pay in fees?" got the same list, with no total. A question about one
amount now gets a written answer. A question that adds amounts up, over one document short enough to
read at once, is answered from the whole document with a total. That answer starts with a heads-up:
the AI model works out totals itself and can miss or miscount items, so check the total against the
sources. Over a longer document it still lists every amount, now with the note that the list cannot
add them up. "Welche Fristen stehen auf der Liste?" and "Welche Kosten zählen zu den Nebenkosten?"
get an answer instead of a list (#649). -
A yes/no question about your documents gets an answer, not a list. Once a document had a deep
index, a question such as "Gilt die Frist für alle Mieter?" ("Does the deadline apply to all
tenants?") was answered at once with every date in the document, because any "all", "every" or
"each" counted as a request for a list. Without a deep index, the same question carried a heads-up
recommending one. Such questions now get a written answer from the document, with no heads-up.
"All" asks for the list when the request starts with it: "Alle Fristen?", "Kannst du mir alle
Fristen nennen?", "What are all the deadlines?". "List every deadline" and "How many payments are
there?" work as before. A list request about several kinds of things now lists the one it names
first: "List all payments and their due dates" lists the payments. "Terms", "names" and "Daten"
no longer pick a list ("List all the medical terms" gets a written answer) (#644). -
"Summarize this document" uses the deep index. With one long document in the chat and its
deep index built, a summary question was still answered from a few matching passages. It now
draws on the deep index's summaries of every section and says it covers the whole document (#642). -
The source markers in an answer about a whole short document now match its sources. When a
skill such as Contract brief read a short document whole, the answer cited markers like [S2],
but its source cards showed none and an evidence review linked nothing to them. The cards now
carry their markers and a review links the answer's markers to them. Answers written before this
update keep their old display (#643). -
A frozen AI model no longer shows as "Running" and healthy. The app now re-checks the model when
a screen asks for its state, and when an answer on it times out. The AI Model screen then says
Not responding and offers a one-click Restart; Diagnostics says it is unhealthy. The app never
restarts it on its own — a slow model looks the same from outside (#599). -
When the AI model crashes on the processor, it comes back by itself. Before, every screen kept
saying it was running and each question failed until you stopped and started it by hand. Now the
app restarts it once and says so; if the same model crashes again, it stays stopped and the app says
that too (#599). -
"The AI model is starting" instead of "No AI model is running". A question or a document task
asked while a model starts — including the file check before loading, which takes minutes on a
freshly copied drive — now says the model is starting and to try again in a moment. Chat's waiting
screen says the same, and its "No AI model is running" message disappears once the model is back
(#599). -
Stopping or switching the AI model while it is answering now keeps the answer so far. Before,
the chat showed "terminated" and everything written so far was lost. Now the answer ends cleanly,
keeps its text, and is labelled Reply stopped. A question the model had not started answering
says "Not answered — the AI model was stopped." If you pressed Try again and then stopped the
model, the previous complete answer stays. While an answer is being written, the AI Model screen
says that stopping or switching the model ends it (#600). -
An answer you stopped, or one cut short by locking the workspace or closing the app, no longer
looks finished afterwards. Before, only the short "Stopped" note in that window said anything;
after a reload, an unlock or a restart the half-written text read like a complete answer, and an
evidence pack built on it said nothing had been cut. Now such an answer keeps the Reply stopped
label, and pointing at it says why. A question that got no answer at all says "Not answered" with
the reason. An evidence pack and its review state that the answer is incomplete and why. If you
pressed Try again and then locked the workspace or closed the app, the previous complete answer
stays (#612). -
Locking, closing the app, or stopping or switching the AI model during a redaction or a document
edit now stops it, and nothing is saved. Before, a redaction fell back to the simple rule-based
detection, dropped the names the model had already found, and asked where to save that copy. After
Lock now the save dialog even opened over the lock screen. A document edit ended with "The
edits could not be completed". Now both say Stopped. Nothing was saved. A save dialog that was
already open stays open, and saving there still saves (#606). -
Redaction now hides the names, addresses and organisations the AI model finds in the whole
document, not only in its last part. On documents with many people in them, the model's answer for
each part of the document was cut off before it was complete, and the app then ignored that part
entirely. The copy kept most names and addresses visible, while the run still reported the redaction
as done: in a test with a 50-entry client register, 8 of 10 names stayed in the saved copy. Documents
with long paragraphs did not get that far: on smaller AI models the redaction or edit failed every
time with "too long". The app now gives each part of the document the room the model actually has,
reads long paragraphs in smaller pieces, and asks again about any part whose answer was cut short.
The same applies to document edits. On a computer without a graphics card, a long document can take
several minutes (#622). -
When the AI model crashes, freezes or fails during a redaction, nothing is saved, and the run says
so. Before, the redaction fell back to the simple rule-based detection and asked where to save
that copy, a few seconds after the crash. The names and addresses the model had already found were
dropped, and only after saving did the run say "no model running", while a model was running. The
same happened every time when parts of a document were too long for a small model's context window.
Now the run ends with "The AI model stopped before the redaction was finished. Nothing was saved."
or says the document is too long for the current model's context window and what to change. Running
a redaction or a document edit while a model is starting, for example right after a crash, now says
the model is starting and to try again in a moment (#620). -
When the AI model crashes in the middle of an answer, the chat says so in words. It used to
show "terminated" or "fetch failed". Now it says "The AI model stopped before the answer was
finished" and what to do next (#600). -
Asking a document question again could delete the previous answer. Choosing "Answer without
it" or "Run with" a skill, then pressing Stop while the documents were still being searched,
removed the previous answer and left nothing in its place. The previous answer now comes back
(#600). -
Apps connected through the local API learn when you stop or switch the model. A request that
was running at that moment ends with "model_not_loaded" (stopped) or "model_starting" (another model
is loading — retry shortly), no longer with the crash error 502 (#600). -
A chat whose AI model never starts on the question now ends after three minutes with a clear
message. When the model's engine accepted a question and then went silent, the chat waited for
five minutes and then showed only "fetch failed". Meanwhile the app counted the model as busy, so
actions such as the engine update were refused. The chat now ends after three minutes with "The
AI model stopped responding", and the model is free again. Apps connected through the local API
get their documentedruntime_unresponsiveerror at the same point. An answer that stops partway
through ends after 30 seconds without new text. With the current AI engine, its keep-alive
messages could postpone that (#594, #598). -
A chat whose AI model gets stuck while it reads a long question now ends with a clear message.
The AI engine keeps signalling that it is alive while it works through a question, even when the
model itself has stopped, so such a chat could wait forever. The app now watches the model's own
progress through the question: when there is none for ten minutes, the chat ends with "The AI
model stopped responding". If the engine goes completely silent instead, the chat ends after two
minutes. A long question on a slow computer still gets the time it needs as long as the model keeps
working through it (#598). -
A search model that stops partway through its reply no longer holds things up for five minutes.
When the search model that prepares documents for questions started a reply and then went silent,
the app waited about five minutes before giving up with "terminated". Until then the rest of an
import waited behind it, a password change or an engine update was refused as busy, and Stop did
not end a document question that was waiting for it. The app now gives up after two minutes, as it
already did when the search model did not reply at all, and Stop ends the question at once (#607). -
Stop and Cancel no longer wait for a model that is still loading. When you stopped a document
question, or cancelled an image analysis or a document translation, while the model it needed was
still starting, the app waited until that model had loaded: up to three minutes if loading got
stuck. Until then a new image analysis was refused as busy, and after a cancelled document
translation the chat stayed blocked. Stop and Cancel now take effect at once. The model goes on
loading so that your next question or job can use it; the image and translation models are
unloaded again after two minutes if nothing uses them (#635, #637). -
A translation whose model gets stuck now ends within minutes, and the next try starts on a fresh
model. When the translation model stopped working partway through a text, the app waited 45
minutes or longer before giving up on that part. In a document translation every later part then
waited the same way, so the job could hold up chat and other document tasks for hours. The app now
watches the model's progress. It gives up on a part after two minutes without a new word, after
ten minutes without progress while the model reads the part, or after two to three minutes of
complete silence. It then restarts the translation model and tries that part once more. A slow
computer still gets the time it needs as long as the model keeps working (#605). -
Building a deep index no longer turns explanations into lists of unrelated items. After "Build
deep index", a question such as "Erkläre was in diesem Arztbrief steht inkl. aller Abkürzungen",
"Liste alle Abkürzungen auf" or "How many pages does the letter have?" was answered at once with
a list of things found in the document. It was not an explanation, and not a list of the
abbreviations either. Questions that ask to explain or describe now get a written answer from
the document, and so do list and count questions about things the deep index does not collect.
List and count questions about dates, amounts, people and organisations, or obligations still
use the deep index. The "Heads-up" that recommends building a deep index now appears only when
the answer really left parts of your documents out. It also no longer promises "a complete
answer": what the deep index adds is a list that covers every section. When a deep index would
not help, a shorter heads-up only says that the answer may miss items (#641). -
"Install voice engine" installs only the voice engine. On a drive with an older AI engine, the
button also tried to download the AI engine again, and while a model was running it then refused
with "The AI engine can't be replaced while a model is running". It now installs only what is
missing; updating the AI engine is the separate, explicit Update (#516). -
A redaction or document edit no longer fails at the very end because of one long suggestion.
When the model suggested a passage longer than the tool accepts, the whole run stopped with "This
tool was given input it cannot accept." after reading the entire document, and every other suggestion
was lost. Such a passage is now skipped and the rest are applied (#583). -
Redaction no longer leaves part of a name visible. When the model suggested both a shorter and
a longer form of the same name (for example "Jane" and "Jane Doe"), the redacted copy could hide
only the first one it got and keep the surname. The same happened when a name and a company shared
a word ("Anna Berg" and "Berg GmbH"). Now every part of the text that any confirmed suggestion
covers is hidden, in the.txtand the Word copy alike, whatever order the suggestions came in.
Redaction stays AI-assisted best-effort: review the copy before sharing it (#580). -
The skill the chat suggests now keeps up with your documents. After you deleted your newest
document and added another, the suggestion could still go by the deleted document and miss the new
one, until some later change to your documents. Archiving a document, or taking one out of a
project and adding another, could leave it out of date the same way. It now follows every change.
A suggestion is only ever an offer: nothing is applied until you pick it (#581). -
Asking to "remove all personal data" now suggests Document Redaction. The chat offered the
skill for "remove personal data" but not for the same request with "all" in it, so a question
like "Remove all personal data for GDPR compliance." got no suggestion (#583). -
German questions about personal data are no longer handed to Document Redaction. A question
that mentioned "personenbezogene Daten", about the GDPR, a privacy notice or a clause in a contract,
was offered Document Redaction. With automatic skills switched on and a PDF selected, it got a count
of the e-mail addresses and phone numbers in the document instead of an answer. Such questions are
now answered normally. Asking to remove the data offers the skill when it is phrased like "Bitte
alle personenbezogenen Daten entfernen." or "Lösche die personenbezogenen Daten …", and with the
skill selected, "lösche" now points to the Redact personal data button like "entferne" does.
Other phrasings get no suggestion, including two that used to get one: "Bitte personenbezogene
Daten löschen." and "eine Version ohne personenbezogene Daten". Pick the skill yourself there (#608). -
Questions that mention sensitive data are no longer handed to Document Redaction. A question
that mentioned "sensitive data" or "sensible Daten", about the GDPR, a privacy policy or a contract,
was offered Document Redaction. With automatic skills switched on and a PDF selected, it got a count
of the e-mail addresses and phone numbers in the document instead of an answer. Such questions are
now answered normally when no skill is selected, in English and German. "Remove the sensitive
data" and "sensible Daten entfernen" still offer the skill. Other requests that name only
sensitive data, such as "Please black out all sensitive data" or "bitte auf sensible Daten
checken", no longer get a suggestion: pick Document Redaction or Share-Safe Review yourself there
(#604). -
What Changed? treats English and German compare requests the same way. With the What Changed?
skill selected and one document in scope (or three or more), "Fasse die Unterschiede zusammen" asked
you to pick exactly two documents. "Summarize the differences" got an answer built from a few passages
of the documents instead, which could describe a second version that was not there. Now both languages
ask for exactly two documents. That covers requests with words such as "differences", "difference",
"changes", "compare", "comparison", "vs" or "old/new version", and German ones such as "Was hat sich
am Vertrag geändert?", "Was ist der Unterschied?" or "die beiden Verträge vergleichen". "Differences"
also suggests the skill now, as "Unterschiede" already did (#582). -
Word documents no longer leave some of their text out of search. Text inside custom XML
(used by some forms and templates) and text moved with Track Changes was skipped when a.docx
was added, so questions about it found nothing. The updated Word reader (mammoth 1.13.0) reads
both, and counts moved text once, at its new place. To pick it up in a document added before
this update, use Re-index on it (#585). -
Sideways and upside-down scans are read the right way up. A page scanned the wrong way round
came out as nonsense text, and the task still ended as done. Text recognition now notices a
page it reads poorly, checks which way up the text is, and reads it again turned upright. This
needs one more OCR file, the page-orientation data (about 4 MB, Apache-2.0): it comes with the
OCR files from now on, and on a drive that already has the language files, the AI Model
screen offers it. To fix a scan read before this update, use Read again (OCR) (#538). -
Photos taken with the phone held sideways are read correctly. Many phones record which way up
a photo is in a form the text recognition ignored, so such a photo was read sideways. It is now
turned upright first (#538). -
A photo of a page now shows that its text came from text recognition. Its preview says so
like a scanned PDF's, it appears under the OCR view in Documents, and its ⋯ menu offers
Read again (OCR) (#574). -
The preview counts only the pages on which text was recognized. A scan with blank pages said
"10 pages" when text was found on 6; it now says "on 6 of 10 pages". Documents read before this
update show the corrected count after the update (#576). -
"Make searchable (OCR)" now reads black-and-white scans. Most office scanners and fax
machines save black-and-white pages in a compact format (CCITT fax or JBIG2), and some scans
use JPEG 2000. The app showed such a page to the text recognition as an empty white page, so
the scan stayed unsearchable with the message that no readable text was found. This has been the
case since text recognition for scans was added. These scans are now recognised like any other.
Run Make searchable (OCR) again on a scan that failed this way (#551). -
Code in answers keeps its line breaks. A code block showed all of its lines run together on
one line, everywhere the app formats text: chat and image answers, translations, document
summaries and the preview of Markdown files. Each line of code now starts on its own line, and
blank lines between them stay. -
The greyed-out microphone now says what is really missing, and its button goes there.
Voice dictation needs two things on the drive: the speech model and the voice engine that runs
it. The note under the greyed-out mic used to blame the speech model every time. On Linux it
kept doing so after the model was installed, when the real gap was the voice engine, which the
app could not install there. The note now names the speech model, the voice engine or both.
Its button says what it does (Get the speech model, Get the voice engine or Set up
voice dictation) and opens the AI Model screen on the speech model itself, not on the
list of installed models. Where the app cannot install the voice engine on this operating
system, the note says that voice dictation is not available there yet and offers no button.
The speech-model card on the AI Model screen says the same. An audio file import that fails
for the same reason now names the missing piece too (#527). -
No more "Open AI Model" buttons. The English label read like the company name "OpenAI".
On Home and in the chat's "no model is running" note the button now says Go to AI Model
(#527). -
Buttons about a missing vision or translation model open the AI Model screen on that model.
Go to AI Model on the Images and Translate screens, and Get the translation model… in a
document's menu, opened the list of installed models, which cannot show a model that is not
installed. They now open the list to download from, filtered to Images or Translation, as the
speech-model button already does (#539). -
A PDF dropped on the Images screen is named as one. The Images screen reads one picture, and
a PDF got the general "That file type isn't supported" message. It now says so, points a scanned
PDF to Make searchable (OCR) under Documents, and offers Go to Documents. The pointer to
Make searchable also stays under the drop area once the vision model is installed (#539). -
A scanned PDF attached in a chat now says what to do. The chat used to show only "This PDF
looks like a scan" and drop the file. It now says to use Make searchable (OCR) in Documents,
offers Go to Documents, and tells you the file then joins the chat by itself. A photo of a
page attached while the drive has no OCR files gets the same kind of note. The note stays with
the chat you attached the file in, even if you switched chats meanwhile (#570). -
A chat whose attached file can't be read no longer answers from all your documents. When
the only file of a chat started by attaching it failed (a scan, a damaged file, an interrupted
import), the chat answered from everything on the drive, including other chats' files. It now
answers from no documents until the file can be read, and lists it as "not readable yet" (#571). -
No more "Try again" on a scanned PDF's row in Documents. Trying again read the same pictures
and failed the same way. The row now offers only Make searchable (OCR) (or Download OCR
files first), and Retry all skips rows a retry can't fix. The Translate screen's note
for a scanned PDF also gets a Go to Documents button (#572, #570). -
Back to your chat from Documents. After a chat's Go to Documents, Documents shows ‹ Back
to chat, which reopens that conversation instead of a new, empty one (#573). -
When your computer can't run the AI engine, the app now says so, names what is missing, and
blames nothing else. Sometimes the engine is on the drive but your operating system refuses
to start it: on Linux when a system library such aslibgomp1is missing, or the system is
older than Ubuntu 22.04 / Debian 12 / Fedora 36; on Windows when Microsoft's Visual C++
runtime is missing, or Smart App Control or an antivirus program blocks the program. Until now the app misreported this:- it switched your graphics card off without telling you, or said your model "could not be
loaded on this computer"; - Translate blamed low memory, and Images suggested another model;
- the GPU check reported "no graphics card";
- documents failed with a raw technical line that showed the drive's folder path.
Now the AI Model screen shows one banner that says what is missing, which package to install
on Linux, and offers Check again for once you have fixed it, with no restart needed. Home,
the chat, Translate, Images, dictation and Documents all point to that banner, and Settings →
Diagnostics has the technical detail. Nothing about it is stored, and drives that were already
marked wrongly (graphics card switched off, failed documents showing a path) are corrected the
next time you unlock. The voice-dictation messages are now shown in German too (#530). - it switched your graphics card off without telling you, or said your model "could not be
-
Home no longer calls the demo mode "running". When a model answers with simulated demo
replies, Home says so (#530).
Security
- The PDF reader that powers "Make searchable (OCR)" now runs fully inside the app's strict
content-security sandbox. When a scanned PDF is prepared for text recognition, the component
that reads it previously ran in a helper that was not covered by the app's content-security
policy. It now runs under the same strict policy as the rest of the app, which blocks any
attempt to reach the network or run generated code. OCR and the text it produces are unchanged. - The app's windows can no longer read other files on your computer. The app's screens used to
be loaded as local files, and in that mode the browser engine behind them could read any file
your user account can open. Nothing in the app did so, but a future bug that let content run code
in a window could have used it. The screens now load from the app's own private address, which
reaches the app's bundled files and nothing else, and the app has given up that extra file access
altogether (#560). - The app can no longer be used to run other programs' code, and it checks its own files.
The HilbertRaum program could be started as a general-purpose script runner, or with a debugging
port, by anything that could already start programs on your computer. Both are switched off now.
On Windows and macOS the app also stops if its main program files have been changed since it
was built. Nothing changes in how you use the app (#562). - Exporting an evidence pack as a PDF no longer puts a temporary copy of the pack next to the
exported file. To print the PDF, the app used to save the pack's readable content as a temporary
file beside the one you chose, and remove it afterwards. If the app was closed mid-export, or
another program held that file open, the copy could stay behind. The app now prints the pack
straight from memory, so no such copy is ever written (#563). - The app no longer contacts Google for a spelling dictionary, or the local network for proxy
settings. Spell-checking has been off in the app for several versions, but the browser engine
inside it still downloaded a spelling dictionary from a Google server when the app started: on
Linux for every language, on Windows for a language Windows cannot spell-check itself. Nothing
from your workspace was sent, but the request told the server your computer's address and
language. On Windows the engine also asked the local network for proxy settings, which a hostile
network could answer. Both requests are gone now. Nothing changes in how you use the app (#567). - Documented: on Linux, the AppImage can run without Chromium's sandbox. On distributions
that restrict unprivileged user namespaces, Ubuntu 24.04 and later among them, the AppImage
starts with Chromium's operating-system sandbox switched off instead of refusing to start, and a
desktop shortcut built from the AppImage's own menu entry always starts it that way. Nothing in
the app changed; Windows and macOS are not affected. Details in the known limitations. - Electron 43.4.0 → 43.7.7, the browser engine the app runs on, with seven weeks of
Chromium security fixes. It clears four high-severity Electron advisories
(GHSA-gr2m-v5gq-v685, GHSA-j84w-jfhq-vhvj, GHSA-9qh4-3jw8-366w, GHSA-qmv3-fv6v-rmhq). Three
concern features HilbertRaum does not use; the fourth needs a separate attack on the app
first. Workspaces open exactly as before, also with an older copy of the app, and the
minimum operating systems are unchanged. The window can take a moment longer to appear at
start, under a tenth of a second on our test PC. That change comes with the security fixes. - Dependency advisories cleared in the build and test tools (
undici,brace-expansion,
fast-uri, and the test runner Vitest), and indompurify, which the dependency graph lists but
the app does not include. None of these packages is part of the app you run, so nothing changes in
use. - The document readers are updated. The Word (DOCX) reader now withstands two kinds of crafted
file, one that could change how it reads a document and one that could stall it. The YAML reader
behind skills limits how often a file can refer back to itself. PDF reading moves to pdf.js 6.3,
which extracts text better from documents that use certain fonts without embedding them. The app
also no longer carries pdf.js's standard font files, which it never used. Workspaces open
exactly as before.
Issues resolved
- #410 — OCR language files are the only drive asset with no in-app install path (follow-up to #59)
- #467 — Tests: zim-client's 8 MiB loopback legs fail with
read ECONNRESETunder CPU load — retries don't rescue them - #512 — Bump the llama.cpp runtime pin b9849 → b11146 (v0.5.0)
- #516 — Tell the user when the drive's AI engine is older than the pinned one (and let them update it)
- #518 — Re-capture the --list-devices and vision SSE fixtures on the current pin (F-40)
- #530 — Linux: a missing system library (e.g. libgomp.so.1) is misreported as a GPU/model failure; document the Linux minimum
- #532 — Offer a reinstall when the AI engine is on the drive but its own files are missing or damaged
- #538 — OCR: sideways and upside-down scanned pages are not detected or rotated
- #539 — Vision model: a downloaded vision model gives no hint where it is used (Images only, no PDFs, chat stays text-only)
- #548 — packaging.test.ts checks lockfile paths, not the packed app.asar, so a nested package can ship past a negation
- #549 — Two tests flake under full-suite load: ocr.test.ts worker-load race (2 s) and zim-client stalled-read retry (4 s)
- #550 — Before taking streamdown 2.6: three tests use an ES2024 API under the ES2022 lib
- #551 — Verify the packaged app on pdf.js 6.3 and the Documents list on react-virtual 3.14.13 (DEP-6 checks not run)
- #554 — kiwix-serve: a cancel that lands during ensureStarted's stop() of the old server is missed
- #555 — Model-slot arbiter: a chat abort right after the builder's handoff is missed until the turn unwinds
- #560 — Serve the renderer from a custom app:// protocol and turn off GrantFileProtocolExtraPrivileges
- #562 — Turn off the RunAsNode, NODE_OPTIONS and --inspect fuses; evaluate asar integrity
- #563 — Print evidence packs from memory instead of a transient HTML file beside the export
- #567 — Packaged app: Chromium fetches the spell-check dictionary on Linux and probes WPAD on Windows
- #570 — Chat: a scanned PDF attached in a chat fails with no pointer to Make searchable (OCR)
- #571 — Chat: a chat started by attaching a file that then fails answers from the whole corpus
- #572 — Documents: a detected scan row offers Try again, which can only fail the same way
- #573 — Chat: no way back to the chat after leaving it for Documents to make an attachment searchable
- #574 — OCR: a photo of a page never gets OCR metadata (no caveat, not in the OCR view, no Read again)
- #575 — OCR: a PDF with some text pages and some scanned pages is never offered OCR
- #576 — OCR preview: "{count} pages" counts pages on which no text was recognized
- #580 — Redaction: a shorter name proposed before a longer one leaves part of the name unmasked
- #581 — Skill suggestions: the scope-signal cache can miss a delete + import and keep stale document titles
- #582 — What-changed: English and German "summarize the differences" route differently at 1 or 3 documents
- #594 — Chat: a llama-server that hangs before sending response headers leaves the chat waiting forever (CB-5 watchdog covers the body only)
- #598 — Chat: the CB-5 prefill budget is wrong on both pins — b9849 cuts real CPU prefills at 120 s, b11146's SSE pings keep a hung prefill alive forever
- #600 — Chat: stopping the model, or a sidecar crash, during an answer shows the raw text "fetch failed"
- #602 — Skills: Document Redaction misses most German removal requests (thin German verb vocabulary)
- #603 — Skills: Share-Safe Review is never suggested for natural German "can I send this?" questions
- #604 — Skills: 'sensible Daten' / 'sensitive data' auto-fire Document Redaction on GDPR, privacy-policy and contract questions
- #605 — Translation: a hung translation sidecar is only cut off by the 45-minute request cap — readCompletionSSE has no idle watchdog
- #606 — Lock and quit do not cancel a running skill run, and the benchmark speed leg has no abort signal
- #607 — Embeddings: the e5 request timeout is cleared at the headers, so a stalled response body is never cut off
- #608 — Skills: a bare 'personenbezogene Daten' suggests Document Redaction, and auto-fires it on GDPR, privacy-notice and contract questions
- #612 — Answers ended by the Stop button, or cut by lock or quit, look finished afterwards — no lasting marker, and evidence packs say nothing was cut
- #613 — A question that got no answer (error, crash, model stopped) cannot be sent again — Try again exists only on answers
- #620 — Redaction: a model crash mid-redaction still falls back to the rule-based floor and offers the copy before saying so
- #622 — Redaction and document edit: the locate windows have no token budget, so long paragraphs overflow a 4,096-token model every time
- #634 — Embeddings: a search-model timeout or start failure reaches the user as raw English text (and a corrupt model's absolute path on every document row)
- #635 — Embeddings: a Stop during the search model's cold start waits for the start (up to 180 s if it hangs)
- #637 — Vision and translation: a cancel during the sidecar's cold start waits for the start (#635's gap)
- #641 — Deep index makes "explain ... inkl. aller X" answers worse: router sends explain questions to the listing engine
- #642 — Chat: the router's tree-summary decision has no consumer — a summary question with a ready deep index is answered top-k
- #643 — Small documents: read the whole document instead of top-k when it fits the context (#641 fix 2)
- #644 — Router: a lookup question about a typed thing with a quantifier gets the full deep-index list ('Gilt die Frist für alle Mieter?')
- #646 — Builds started from a lower-case drive letter (c:\…) bundle ~41 main modules twice; main-side messages then come out in English on a German system
- #649 — Router: a 'how much' lookup about a typed thing gets the full deep-index list ('How much is the monthly fee?')
What's Changed
- docs: Linux AppImage launch steps and start-up troubleshooting by @humaniser in #528
- fix(dictation): name the missing piece and link to the speech model (#527) by @comilionas in #529
- fix(engine): name an engine the OS refuses to start — never the GPU or the model (#530) by @comilionas in #531
- docs: point the present-but-broken engine reinstall at its own issue #532 by @comilionas in #533
- fix(deps): Electron 43.4.0 → 43.7.7, clearing the four high Dependabot alerts (DEP-5) by @comilionas in #534
- fix(deps): DEP-6 PR A, lockfile-only patch batch (undici, brace-expansion, fast-uri, dompurify) by @comilionas in #535
- fix(deps): DEP-6 PR B, vitest 3.2.6 -> 4.1.11 (alerts 91, 92, 94) by @comilionas in #536
- chore(deps): DEP-6 PR C, add .github/dependabot.yml (D-4) by @comilionas in #537
- fix(deps): Dependabot versioning-strategy lockfile-only (DEP-6, after #537) by @comilionas in #542
- test: vision-security waits for each image's history write (vitest 4 teardown flake) by @comilionas in #546
- docs(deps): correct the lockfile-only claim in dependabot.yml (DEP-6, after #542) by @comilionas in #544
- chore(deps-dev): Bump the development group across 1 directory with 5 updates by @dependabot[bot] in #540
- fix(deps): Dependabot's production group (#545) with its review fixes — pdf.js 6.3 and six more by @comilionas in #547
- chore(deps): Bump the production group across 1 directory with 7 updates by @dependabot[bot] in #545
- fix: the DEP-6 residuals — packed-layout gates (#548), load races (#549), lib ES2024 (#550), #551 checks by @comilionas in #552
- chore(deps): streamdown 2.6 — mermaid leaves the tree; Markdown layout fixed without Tailwind (stacked on #552) by @comilionas in #553
- fix: two missed aborts across an await (#554, #555) by @comilionas in #556
- fix(ocr): ship pdf.js's scan decoders so black-and-white scans OCR (#551) by @comilionas in #557
- fix(ocr): run pdf.js inside the rasterizer page, under its CSP by @comilionas in #558
- docs(architecture): correct the OCR in-page record's size figure, add timing + mutation checks by @comilionas in #559
- Serve the app's pages from a custom scheme and turn off GrantFileProtocolExtraPrivileges by @comilionas in #561
- Packaged build: turn off RunAsNode and --inspect, turn on asar integrity (#562) by @comilionas in #565
- Evidence-pack PDF export: print from memory instead of a temporary file beside the export (#563) by @comilionas in #566
- Platform verification of #560/#562: Linux AppImage measured, macOS build inspected, owner-run Mac check by @comilionas in #568
- Packaged app: no spell-check dictionary download and no WPAD lookup from the browser engine (#567) by @comilionas in #569
- fix(models, images): say where each automatic model is used; land missing-model buttons on the model (#539) by @comilionas in #577
- fix(chat, documents): a file attached in a chat that can't be read yet (#570, #571, #572, #573) by @comilionas in #578
- OCR: read sideways pages upright, say when it was unsure, give photos OCR metadata, read a text PDF's scanned pages (#538, #574, #575, #576) by @comilionas in #579
- chore(deps-dev): Bump the development group with 2 updates by @dependabot[bot] in #584
- test(skills): prune and repair the skills test surface (campaign part 1 of 3) by @comilionas in #586
- Install the AI engine again when its own files are damaged (#532) by @comilionas in #587
- chore(deps): mammoth 1.12.3 → 1.13.0, with regenerated notices (replaces #585) by @comilionas in #589
- fix(redaction): mask the union of overlapping proposals, never just the first (#580) by @comilionas in #590
- test(skills): run the trigger eval through the production suggestion and auto-fire paths (campaign part 2 of 3) by @comilionas in #588
- test(skills): move the copy-pasted skills test setup into shared helpers (campaign part 3 of 3) by @comilionas in #591
- test(skills): finish the test campaign's open steps by @comilionas in #592
- feat(engine): report engines older than the app's pin and update them in place (#516) by @comilionas in #593
- fix(skills): suggest Document Redaction for "remove all personal data" (#583) by @comilionas in #595
- test(skills): cover the skills IPC, routing, echo-strip and read-failure gaps (#583) by @comilionas in #596
- fix(skills): drop over-long locate proposals; share the locate windows (#583) by @comilionas in #597
- fix(chat): bound the wait for llama-server's response headers (#594) by @comilionas in #601
- fix(skills): #583 findings follow-ups — grounded-data echo strip, wave3 paths, registry scan wording, test-command docs by @comilionas in #609
- fix(skills): stop offering Redaction on a bare "personenbezogene Daten" (#608) by @comilionas in #610
- docs(skills): record why German removal requests get no keyword fix (#602) by @comilionas in #614
- docs(skills): record why German "can I send this?" questions get no keyword fix (#603) by @comilionas in #615
- fix(skills): stop offering skills on a bare "sensitive data" / "sensible Daten" (#604) by @comilionas in #616
- fix(chat): end a prefill wedged behind SSE pings; headers get 180 s (#598) by @comilionas in #611
- feat(runtime): a frozen or crashed chat model no longer reads "running, healthy" (#599) by @comilionas in #618
- fix(chat): a model stop or switch ends the answer cleanly; a crash speaks in words (#600) by @comilionas in #617
- fix(chat): the Stop button, lock and quit leave a lasting "Reply stopped" marker (#612) by @comilionas in #619
- fix(teardown): lock, quit and a model stop cancel skill runs and the benchmark speed leg first (#606) by @comilionas in #621
- fix(redaction): a model that fails during the locate pass fails the run, with nothing saved (#620) by @comilionas in #623
- fix(skills): the redaction and edit locate windows and replies fit the model's context (#622) by @comilionas in #624
- feat(chat): "Send again" for a question that got no answer (#613) by @comilionas in #625
- fix(skills): the skill suggestion follows every change to the documents in scope (#581) by @comilionas in #626
- fix(skills): What-changed routes English and German compare requests alike (#582) by @comilionas in #630
- fix(embeddings): the embed request's deadline and Stop cover the response body (#607) by @comilionas in #633
- fix(embeddings): a search-model failure reaches the user as words, chosen by kind (#634) by @comilionas in #636
- fix(embeddings): a Stop ends a document question's wait for a cold start, not the start (#635) by @comilionas in #638
- fix(vision,translation): a cancel ends a job's wait for a cold start, not the start (#637) by @comilionas in #639
- fix(translation): a hung translation sidecar ends within minutes and is restarted (#605) by @comilionas in #640
- fix(rag): explain requests and untyped list asks get a written answer; the deep-index hint shows only when the answer left sections out (#641) by @comilionas in #645
- fix(build): a build started from a lower-case drive letter bundles each module once (#646) by @comilionas in #647
- Small documents are read whole; a summary ask with a ready deep index reads the tree (#643, #642) by @comilionas in #648
- A quantifier asks for a list only at the head of a request; a yes/no question gets an answer (#644) by @comilionas in #650
- fix(rag): "how much" is a lookup or a sum; a list asked about and "zählt als" ask for no list (#649) by @comilionas in #651
- release: v0.1.63 — version bump by @comilionas in #652
New Contributors
- @dependabot[bot] made their first contribution in #540
Full Changelog: v0.1.62...v0.1.63