A System-Only or AI-Only run is a partial result, not the NOMAD Score
(which is the full-benchmark composite). Two problems addressed:
1. Scoring bug: AI-only runs were NOT renormalized. _calculateNomadScore
always added the system weights (0.60) to the denominator even for an
AI-only run (default-zero system scores), so an excellent AI-only run
scored ~39.8 -- scaled against the full NOMAD 100 where AI is only 40%
-- while system-only already renormalized correctly. Fix: pass
systemScores only when the system benchmarks actually ran, so AI-only
renormalizes to its own 0-100 (39.8 -> ~99.7). Full and System-only
scores are unchanged.
2. Presentation: partial runs were shown with the full "NOMAD Score"
label + big gauge, outweighing the small "Partial" notice. Now partial
runs are relabelled "System Score" / "AI Score" with a PARTIAL badge,
a muted (neutral) gauge + number, and a "run a Full Benchmark for your
NOMAD Score" CTA -- applied to both the persistent score section and
the Phase 3 ScoreReveal via a shared getScoreDisplay() helper. Adds a
`muted` prop to CircularGauge.
Browser-validated on NOMAD3: AI-only now shows "AI Score" + PARTIAL,
muted gauge, 99.7 (was 39.8); Full still shows "NOMAD Score" green.
Implements the display-layer fix from the Score v2 red-team's W2; the
AI-ceiling saturation (W1) remains v2 work.
Part of #1082.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 3 of the live benchmark run experience.
- ScoreReveal: replaces the abrupt run-view unmount with a deliberate
end-of-run "REPORT" card -- animated NOMAD score gauge + odometer
count-up number, sub-score gauges cascading in, Continue button +
5s auto-dismiss. Takes score scale as a prop so it survives Score v2.
- GPU-util overlay (NVIDIA): during the AI stage, a ~1Hz nvidia-smi poll
inside the Ollama container feeds live GPU utilization + VRAM into the
telemetry frames; shown in the AI hero, hidden when absent (AMD/none).
Poller is side-effect-only and cleared in a finally -- scored numbers
unchanged.
- Disk polish: reset in-test buffers on stage transition so the write
stage no longer briefly shows the carried disk-read value.
Browser-validated on NOMAD3 (RTX 5060): GPU overlay live (1% util,
0.2/8.0 GB VRAM during model load); reveal cascade + count-up; AI-only
score 39.8 (scored path unchanged).
Part of #1082 (tracker stays open).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 2 of the live benchmark run experience. Streams sysbench
--report-interval=1 interim lines over a Docker attach so the run view
shows authoritative in-test CPU events/sec and disk read/write MiB/s
(overlaying the Phase 1 host-proxy disk numbers), and fills a
"results so far" strip as each stage completes.
- _runSysbenchCommandStreaming: attaches to the container output for
live onLine callbacks, but returns the authoritative output via
container.logs() after exit -- byte-identical to _runSysbenchCommand,
so the SCORED numbers are unchanged (attach 'data' can flush after
container.wait() resolves, which would truncate the final report).
- CPU + disk stages stream interim eps: / reads:/writes: MiB/s into
setStageMetric; memory stays non-streaming (too fast to sample).
- _emitPartialResult broadcasts each finished stage's raw result on the
progress channel; useBenchmarkRun accumulates them; ResultsSoFar
renders the chip strip.
- Frontend: live CPU ev/s readout+sparkline, disk hero switches to
"Benchmark throughput" when in-test numbers arrive.
Browser-validated on NOMAD3 (System-Only): CPU 6536 ev/s live, disk
10399 MB/s benchmark throughput, results strip, final score 68.3.
Part of #1082 (tracker stays open for Phase 3).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 1 of the live benchmark-run experience. Replaces the opaque (and
in sync mode, simulated) progress bar with a real-time run view driven by
actual host telemetry.
- Async run path: the UI now dispatches to the queue worker and keys off
SSE instead of faking stage progress with client-side timers.
- New BenchmarkTelemetrySampler broadcasts per-core CPU load, CPU temp
(best-effort, hidden when unavailable), and disk MB/s at 1 Hz over a new
benchmark-telemetry SSE channel. Runs in the orchestration process, never
the sysbench container, so it cannot affect scores.
- BenchmarkProgress carries the ordered stage plan + index so the frontend
renders a live stage rail.
- AI benchmark streams /api/generate for live tokens/sec and true TTFT; the
scored numbers still come from Ollama's authoritative final eval fields.
- Frontend: useBenchmarkRun hook owns both subscriptions; self-contained SVG
components (StageRail, CoreGrid, Sparkline, LiveReadout) + BenchmarkRunView,
styled in the desert palette. No chart library added.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rebuilt on top of dev's RFC #883 state-machine UI rather than the now-defunct
StoredFile shape:
- Extend StoredFileInfo with fileName/size/uploadedAt/isUserUpload
- Populate metadata from on-disk stats in RagService.getStoredFiles
- Add fileSourceSchema validator + getFileContent/downloadFile endpoints
scoped to the uploads directory only (tighter than the original PR — matches
docs_service traversal pattern)
- KnowledgeBaseModal: sortable Size and Uploaded columns; View/Download
buttons on upload-bucket rows; new FileViewerModal for in-browser text
preview. Bucket grouping preserved — sort applies within each bucket.
- Use formatBytes from ~/lib/util rather than redefining
BullMQ instantiates a fresh ioredis client per Queue/Worker when handed a
plain {host, port} config object, and under sustained ZIM ingestion the
embed pipeline leaked ~1 client/sec until Redis maxclients was exhausted.
Pass a single shared ioredis instance (maxRetriesPerRequest: null, as
required by BullMQ) so all queues and workers reuse one client pool.
Workers still duplicate the connection once for their blocking client,
which is expected and bounded.
Closes#885
* feat: replace legacy Kolibri image default with latest v19 image
* feat(supply-depot): add content migration instructions for Edu Platform Gen 1 to 2
Adds the MeshCore web client to the Supply Depot catalog (host port 8500),
alongside the existing Meshtastic apps. Uses aXistem's prebuilt image of Liam
Cottle's MeshCore client (MeshCore is a sibling LoRa mesh project to Meshtastic).
The image is stock nginx serving a static Flutter build over HTTP, but the
client reaches radios via Web Bluetooth / Web Serial, which browsers only allow
from a secure (HTTPS) context. So we serve it over HTTPS: a new preinstall hook
generates a self-signed cert + a small SSL nginx config into storage/meshcore-web,
both bind-mounted into the container (the config over the image's default.conf),
publishing 443. Same one-time browser-warning approach as Vaultwarden, whose
openssl cert generation is refactored into a shared _ensureSelfSignedCert helper.
Also adds a NOMAD-specific docs section + Manage>Docs anchor, and registers the
IconAntenna icon. Meshtastic Web left unchanged.
Validated on NOMAD3 (v1.33.0-rc.1): the image + SSL config + self-signed cert
serves the MeshCore Flutter app over HTTPS 200 with working SPA fallback.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Curated catalog apps could be installed, stopped, and force-reinstalled,
but never removed — the only path off a device was manual docker + DB
surgery. Custom apps already had delete; this adds the equivalent for
curated apps.
POST /api/system/services/uninstall stops and removes the app's
container (optionally its image, same best-effort semantics as custom
app delete) and flips the record back to not-installed so the card
returns to the available catalog. Host bind-mount data is deliberately
left on disk, so a later reinstall picks the app back up where it left
off — unlike force-reinstall, which clears volumes.
Guards: custom apps are rejected (use delete), dependency services are
rejected, and uninstalling a not-installed app is a 409.
UI: installed curated cards get an Uninstall action in the card menu,
with a confirm modal that explains data is preserved and offers the
same remove-image checkbox as custom app delete.
Bring the GitHub-facing README in line with the v1.33 feature set:
- Add Supply Depot (one-click app catalog + bring-your-own custom
Docker containers) and Automatic Updates to the "Built-in
capabilities" list and the "What's Included" table.
- Fix the Offline Maps row, which claimed "navigation" — NOMAD maps
are download/view only, with no routing. Reword to "offline viewing
and search."
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Refresh the in-app Markdoc docs for the v1.33 feature set:
- Repoint dead /settings/apps links to the Supply Depot (/supply-depot)
across home, getting-started, and faq; reword "Apps page" / "Settings
-> Apps" to "Supply Depot". The old /apps route now redirects to the
Supply Depot.
- Expand supply-depot-apps.md with a "Managing your apps" section (Docs/
Edit/Logs/Stats/Update/Remove, version + update-available visibility,
custom launch URLs, per-app auto-update toggle) and a "Bringing your
own app" section for custom Docker containers.
- Add a new "Updates" doc (updates.md) covering the auto-update trilogy
(core/apps/content), manual updates, and the Early Access channel;
wire it into DOC_ORDER and cross-link from home, getting-started, faq.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Only Wikipedia had version cleanup; every other curated map and non-Wikipedia
ZIM left its prior version on disk when a newer one installed, so users silently
accumulated orphaned content (potentially hundreds of GB). (#634)
The install paths already record each resource via InstalledResource
{resource_id, resource_type, version, file_path}, so the authoritative old-file
path for a resource is known. On install of a new version we now capture the
prior row before updateOrCreate repoints it, then delete the old file — gated
behind a pure, fully unit-tested decision function with strict safety rails:
- tracked-only: requires a prior InstalledResource row for the same
resource_id, so sideloaded/untracked files are never touched
- genuine replacement: old and new file paths must differ
- new-file-verified: the new file must be confirmed on disk first
- strictly-newer: a re-install or downgrade can't wipe a newer file
- within-storage-dir: the old path must resolve under the content store
ZIM cleanup deletes the old file directly (NOT via this.delete(), which would
drop the InstalledResource row by resource_id that updateOrCreate just
repointed) and rebuilds the Kiwix library only if a file was actually removed,
so its XML never references a deleted ZIM. Maps need no library step. Wikipedia
keeps its own existing cleanup path. All deletions are best-effort and logged;
a failure never breaks the install.
Closes#634
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`getChatSuggestions` previously picked the largest installed model by file
size, on the assumption that bigger models give better suggestions. This
is unsafe: if any installed model exceeds available VRAM (e.g.
llama3.1:405b on a 96 GB GPU), Ollama spends minutes trying to load it
and the request 500s — making the chat page unusable for anyone who
happens to keep a flagship-sized model on disk.
Chat suggestions are short prompts that don't benefit from a flagship
model anyway. Prefer the user's selected `chat.lastModel` when set, and
fall back to the smallest installed model otherwise. `OllamaService.getModels()`
already excludes embedders, so the fallback always picks a chat model.
Two card tweaks for the update workflow:
- Show the installed image tag next to the powered-by name (e.g. "Kiwix ·
3.7.0"), so the running version is visible at a glance. Only rendered for
installed apps; falls back to just the version when powered_by is unset.
- Change the "Update available" pill from a muted light-green tint to a solid
desert-orange fill with white text, so an available update actually draws
the eye instead of blending into the card.
listTags() follows the registry's Link-header pagination, but the next-page
URL is relative per the OCI/Docker registry spec (e.g.
"/v2/ollama/ollama/tags/list?last=0.9.3-rc5&n=1000"). The code assigned that
raw relative path straight back to `url` and re-fetched it, so fetch() threw
"Failed to parse URL from /v2/...". Any image repo with more than 1000 tags
paginates, so the entire tag list — and therefore the update check — failed
silently for ollama/ollama and filebrowser/filebrowser.
That's the root cause of #945 ("won't update past 0.24.0"): the Ollama
update check never completed, so no newer version was ever offered.
Resolve the next-page URL against the registry origin with
new URL(next, `https://${registry}`), which also passes absolute next-URLs
through unchanged for registries that return those.
Closes#945
The maps page reset to the default US-wide view on every refresh because
initialViewState was hardcoded. Save the position and zoom to localStorage
on each move-end (key nomad:map-view, matching the existing scale-unit
pattern) and restore it at mount: saved view → default. The restore is
bounds-checked so a corrupt value falls through to the default.
Replaces #815, whose branch had drifted far out of scope (56 files of
stale-base/merged-commit noise plus unrelated map-feature WIP). This is
just the persist-view improvement, ported cleanly onto current dev. The
null-island and coordinate-search parts of #815 targeted URL-param code
that never landed on dev, so they don't apply here.
The map style names each source by its date-stripped region (both
"washington.pmtiles" and "washington_2025-12.pmtiles" -> "washington").
When an old and new copy of the same region are both on disk, the style
emitted two sources with the same key and duplicate layer ids, which
MapLibre rejects outright -- blanking the ENTIRE map, not just that region.
Old copies linger when a newer curated version installs (#634), so a user
who updates maps can silently lose all map rendering until the stale file
is removed by hand.
generateSourcesArray() now keeps only the newest file per region: a dated
build beats an undated legacy file, and between two dated builds the later
YYYY-MM wins. The skipped duplicate is logged. The style stays valid even
when stale files are present.
Complements #981, which removes superseded curated files on install. This
is the runtime safety net that also recovers installs already in the broken
state (which a cleanup-on-install alone can't reach).
Refs #634
Dense source content produces chunks that exceed the embedding model's
context window (nomic-embed-text:v1.5 defaults to 2048 tokens). Two paths
hit this even after the prior pre-cap:
- Older Ollama (e.g. 0.18.1, #944) ignores the num_ctx=8192 we send on
/api/embed, so it stays at the model's 2048 default.
- The OpenAI-compat /v1/embeddings fallback didn't pass num_ctx/truncate
at all, so any Ollama drops to 2048 whenever it lands on the fallback.
When a chunk overflowed, the 400 was swallowed and the chunk was silently
dropped from Qdrant. Worse, the failure propagated to EmbedFileJob, which
re-embeds the entire file on each of its 30 BullMQ attempts — the "endless
queue loop" / "api/embed for weeks" / pegged GPU reported in #944/#959.
Fix:
- OllamaService.embed(): on a context-length error, retry once with an
aggressive 2048-safe cap (EMBED_CONTEXT_SAFE_CHARS = 2000) so the chunk
is embedded (start-of-chunk) instead of dropped. Native-path context
errors now bubble to this retry instead of falling through to the
smaller-context fallback. Split the native+fallback attempt into
_embedWithFallback().
- Pass truncate/num_ctx on the /v1/embeddings fallback too (Ollama's
OpenAI-compat shim forwards them).
- EmbedFileJob: classify "input length exceeds context length" as an
UnrecoverableError so one permanently-oversized chunk can't trigger 30
full-file re-embeds.
- Add OllamaService.isContextLengthError() shared by both.
Graceful degradation: a truncated chunk loses its tail but is kept in the
index, which is strictly better than today's silent drop + retry storm.
Refs #881. Supersedes the #369/#670 symptom closures that never fixed the
fallback path.
reconcileFromFilesystem() skipped every ZIM whose filename starts with
`wikipedia_en_`, on the assumption that all such files are managed by the
WikipediaSelection model. But curated category tiers ship Wikipedia-themed
ZIMs (e.g. Medicine → Comprehensive includes `wikipedia_en_medicine_maxi`),
so those files were skipped during reconcile and their InstalledResource
rows got wiped on every restart — silently downgrading the detected tier.
Skip only the single file actually tracked by WikipediaSelection, matched
by exact filename instead of the `wikipedia_en_` prefix.
Reimplemented in-house from @johno10661's PR #774 (which was trapped on a
stale base); credit to them for the diagnosis and fix.
Closes#774
A multi-GB service update (e.g. nomad_ollama pulling ~6.5 GB) left the
Update button clickable with no feedback, so users clicked again thinking
it was stuck. The second click raced a concurrent updateContainer run into
Docker 304/400 errors (stop/rename on a container the first run had already
moved). The backend lock was in-memory only and never written to the DB, so
nothing durable signaled "update in progress" to the UI, and a page reload
mid-pull re-enabled the button.
Backend (docker_service.updateContainer):
- Set installation_status='installing' when the update starts and reset it
to 'idle' in a finally on every exit path. This mirrors the install path,
survives a page reload, and is visible to other tabs/clients.
- Reject a second update with a clear message when installation_status is
already 'installing', instead of letting it race into Docker errors.
Frontend (settings/apps.tsx):
- Track in-flight updates per service. Seed optimistically on click and
reconcile with the durable installation_status from the server.
- Disable the per-service Update button and show "Updating..." while in
flight. Drop the fullscreen spinner for updates so the table and the
activity feed (live pull/stop/start progress) stay visible.
Closes#931
Chris Sherwood
cosmistack-bot
jakeaturner
Jake Turner
Benjamin Smith
Henry Estela
Lorenzo Galassi
johno10661
akashsalan