AI page: Quick Check vs Calibrate, host-specific verdicts, colour scheme
Documents that Calibrate is a superset of Quick Check rather than a
separate step people have to run afterwards, since two buttons side by
side read as alternatives. Explains why each candidate batch size has to
pass twice -- silent omission is intermittent, and one clean run is a
single sample of something that fails only some of the time.
Adds when Quick Check alone is the right call: auditioning models,
re-checking after a pull on unchanged hardware, a temporary host, or very
few indexers.
Adds the four-state colour scheme and the Host column, and records that a
verdict belongs to a (host, model) pair -- with the measured 436s vs 237s
difference for the identical model and batch size on two machines.
Corrects the stale claim that 8k context is the practical minimum. That
predates the batching work and contradicted the "Small context windows"
section further down the same page.
Documents the notification narration fallback, after a security-log model
returned a MITRE ATT&CK object that went out verbatim as a push
notification.
docs: model updates and context-window limits
The AI page gains two sections.
Model updates: what the badge means in each of its two states, that the check
never downloads anything to find out, and that auto_update_check is the one
thing that leaves your network.
Context windows: leads with the finding rather than the setting. Fitting the
window is not the whole story -- 25 indexers score, 30 makes the model echo
the input back with the window half full, and 37 in a window twice as large
returns well-formed JSON that silently leaves five out. That last one is what
someone choosing a model needs to know, because it looks exactly like
success.
docs: document the AI master switch
Follows 4916959 / 44473b5, which added an Enable AI features toggle to
Settings -> AI and made it the single control for every AI feature.
AI-and-Model-Validation
New "The master switch" section up front: what stops when it is off, that
no Ollama request is made at all, and that turning it off erases nothing --
URL, model and system prompt stay on disk so switching back on restores
what you had. Notes the upgrade behaviour: an install that already had a
working URL stays enabled, only new installs start off.
Removed "Leave the Ollama URL blank and everything else works unchanged" --
that advice predates the switch and is now the wrong way to do it.
Configuration Added ollama.enabled to the example and the key table,
including the upgrade behaviour. Reworded the Ollama intro,
which said "when configured" and now also means "when
enabled".
FAQ "Do I need Ollama" now says it ships off and names the
switch, instead of telling readers to blank the URL.
Home AI scoring is called out as off by default.
Web-UI The AI pane row leads with the master switch.
Verified: no stale "leave the URL blank" phrasing remains anywhere, all 86
internal links across 17 pages resolve, every file is valid UTF-8.
docs: bring the wiki up to date and add navigation
Structural
_Sidebar.md persistent navigation on every page, grouped getting-started /
how-it-works / using-it / help. GitHub renders this alongside
every page; without it readers had to return to Home to get
anywhere, and Home's hand-maintained index had already drifted
out of sync with the actual filenames.
_Footer.md repo / releases / issues links on every page.
New pages
Quarantine-and-Severity severity grading, the three bands, the review
queue and its three outcomes, timeouts. This was
previously undocumented and would otherwise have
been split across three pages.
AI-and-Model-Validation Ollama setup, the three validation tests, the
gate and the forced state, context requirements.
AI content had been scattered across
Configuration, Prowlarr scoring and Notifications.
API /api/health, /api/status, /api/scan, /api/logs.
The JSON API existed and was entirely
undocumented; Home did not mention it at all.
Fixed
Home [Docker](Docker) and [CLI](CLI) were broken -- no such
pages; the files are Installation and CLI-Usage. Rewritten
with a task-ordered getting-started path, corrected
Pushover -> Apprise (stale since v1.5), and a features list
that reflects quarantine, severity and model validation.
Detection The action sequence went straight from match to delete. It
now documents grading, the three bands, the evidence trail
and correlation IDs, and the client-delete failure path.
Also corrected Pushover -> Apprise.
Web-UI Added Quarantine and Settings -> AI. Corrected the Settings
pane list, which named panes that do not exist and omitted
several that do. Stats is a tab in the Indexers hub, not a
page, since v1.6.0.
Configuration Added the remediation block. Completed the ollama table --
system_prompt and update_check_hours were missing -- and
noted the measured context requirement.
Notifications Added the quarantine event and its fallback to `action`.
Noted that the Ollama toggles are gated on AI being set up.
FAQ "Will it delete good torrents" now covers quarantine. Added
two questions: whether AI is required (it is not) and how
to check a model before relying on it.
Troubleshooting Deleted-good-torrent now explains how to trace it via the
inspection record and three ways to prevent a repeat. Added
stuck-in-quarantine, nothing-ever-quarantined, and Ollama
returning nonsense.
CLI-Usage --daemon also sweeps quarantine timeouts.
Installation Re-encoded from CP1252 to UTF-8. It held 22 non-ASCII bytes
that GitHub renders as mojibake.
Verified: all 86 internal links across 17 pages resolve, and every page is
reachable from the sidebar.