Repository navigation
fix(docs): stop an English household getting German summaries - #86
Merged
Merged
Conversation
Reported from Discord: whisper transcribes correctly, then the summary
comes back in German on an install set up in English. The German was
ours. Both classify prompts carried hardcoded German worked examples and
shipped them to every install regardless of its language.
A small model does not weigh "use the content's language" against a
concrete example; it copies the example. Reproduced against a 9B — the
RAM-tier default famstack itself picks for 16-32 GB — with an empty vault
and English input:
note about brakes -> tags ['bremsen', 'auto-wartung']
note about camping -> "Vorräte und Zeltstangen prüfen für
Campingurlaub", summary in German
Which is why it looked intermittent rather than broken: the drift fires
where the household's subject matter collides with one of our examples.
Cars, camping, insurance, school, tax. Most of what a family server is
for.
Examples are now chosen by the household's language, and the rule names
no language at all, because naming one invites it. German households keep
their German examples, which is the only place those ever helped.
Verified on the same 9B that content language is still followed in both
directions, including the case removing the examples put at risk: German
content on an English install still gets a German title and German tags.
arthware-dev
added a commit
that referenced
this pull request
Oct 6, 2026
Image captures in a German household were filed with English titles,
summaries and tags. The classify prompts told the model to write in
"the content's own language". A pasted photo has no words, so the model
followed the language of the prompt, which is English. The same rule
kept a Spanish invoice in Spanish, although the family reads the wiki
compiled from these entries in their own language.
The archivist now writes title, summary, facts, action items and tags
in the household language (`[core] language`), whatever language the
source is in:
- the prose is translated; names, amounts, numbers, addresses and
established technical terms ("Kubernetes") stay as printed
- the document itself, its PDF and OCR text, is not changed
- the prompt names only the household language, so an English
household is never shown German
- language names come from one shared table, which the diary uses too
German content filed on an English install is now summarized in
English. Entries filed before this change keep their language.
Fixes: 2edcce3 ("fix(docs): stop an English household getting German summaries (#86)")
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
From a Discord report: whisper transcribes correctly, then the summary
comes back in German on an install set up in English.
The German was ours. Both classify prompts carried hardcoded German
worked examples and shipped them to every install regardless of language.
A small model does not weigh "use the content's language" against a
concrete example; it copies the example.
Reproduced on a 9B — the RAM-tier default famstack picks for 16-32 GB —
with an empty vault and English input:
That also explains why it looked intermittent: the drift fires where the
household's subject matter collides with one of our examples.
Examples are now selected by household language and the rule names none.
German installs keep their German examples, which is the only place those
ever helped.
Verified on the same 9B that content language is still followed both
ways, including the case removing the examples put at risk: German
content on an English install still gets German output.