The post-rerank relevance floor rejects noise but passes coherent
off-topic questions, and it gets worse as the library grows. Adding
Gutenberg prose, the Python tutorial and MDN pages to the eval corpus
took the off-topic leak rate from 22% to 67% at the 0.62 floor.
Score cutoff does not fix it. The weakest correct top hit scores 0.666
while "write me a haiku about autumn leaves" pulls Walden at 0.705.
A cosine floor and a top-vs-median contrast test both cost recall
faster than they cut leaks.
- Add rag.relevanceCheck (default off). After the floor, the tasks
model, or the chat model when none is set, gives one grammar-constrained
on_topic verdict for the whole retrieved set. Off topic means no
context and no citations. Any failure keeps the chunks unjudged.
- Add a Settings > Models toggle recommending an 8B+ tasks model.
On llama3:8b the check cuts off-topic leaks from 67% to 6% for about
5 points of core recall@5 and ~0.6s per turn. On 3B models it rejects
most valid questions, which is why it is opt-in.
- Add 6 distractor documents and an off_topic suite (18 unanswerable
questions, 7 distractor controls) to the eval corpus.
- Report per-tag leak rates and leaked cases with top scores in
eval:retrieval. Add --judge-model to measure the check and --dump to
export pre-floor candidate pools for offline gate simulation.
- Pin the check off in eval:generation so a local setting cannot move
scores.
- Record the distractor sweep on RAG_MIN_FINAL_SCORE. The 0.62 default
stays: 0.65 halves leaks at no measured recall cost but leaves only
0.016 of margin.
The corpus fingerprint changes to b4ebe5dce8699f5d. Retrieval reports
from before this commit are not comparable, and a new baseline is
committed.
Refs #1341
Chat sent num_ctx, num_predict and nothing else, so every turn decoded at
whatever the backend picked: temperature 0.8, top_p 0.9, no min_p. On a
heavily quantized 1-3B model that is the wandering, repetitive output people
read as "the local model is bad".
Adds `ai.responseStyle` with four profiles. min_p is the lever, not
temperature: it floors relative to the top token, so it truncates hard on a
confident distribution and barely at all on a flat one, which is where small
models start to wander. top_p stays at 1.0 rather than stacked on top of it.
repeat_penalty moves with the style, because a grounded answer legitimately
repeats a part number inside the 64-token window: `focused` relaxes it to
1.05, `creative` tightens it to 1.15.
`auto` is the only style that defers to the model. Plenty of library models
ship author-tuned samplers (qwen3 sets 0.6/0.95/20), so those win key by key
and NOMAD fills only the gap, which in practice is min_p. A style the user
picked by name overrides them, or it would silently do nothing on half
their models.
The compat transport drops min_p, so a profile carries a separate compatTopP.
Passing top_p 1.0 to a backend with no min_p behind it would sample the raw
distribution, which is worse than the 0.9 it replaced.
Resolved in the controller, so the grammar-constrained task calls keep their
temperature 0. An explicit temperature still outranks a profile's, which is
what keeps the eval harness deterministic.
`eval:generation --style` is new and load-bearing: temperature 0 makes min_p,
top_p and top_k inert, so a default run cannot see this change at all.
`--style=off` sends no samplers and no temperature, and is the control.
Verified against Ollama 0.32.9, min_p accepted. 22 unit tests.
There was one cutoff, Qdrant's score_threshold at 0.3, and nomic scores
unrelated text well above it, so every question retrieved *something*
(nonEmptyRateOnRefusal 1.0) and rag_context wasted four lines asking a 3B
model to "silently judge" relevance. That is a classification a number
should decide, not a micro model.
Adds `RAG_MIN_FINAL_SCORE`, a floor on the post-rerank score, applied before
source diversity. The floor judges relevance, the 0.85^n penalty judges
redundancy. Nothing clears it and nothing is injected. `score_threshold` stays
at 0.3 as the candidate net. Exposed as `rag.minRelevance` with a module-
scoped TTL cache invalidated from `updateSetting`.
The floor is calibrated and the harness was describing the wrong number:
aggregate() bucketed semanticScore, so its percentiles measured raw cosine
while the new knob filters the reranked score. Both axes now, and the gap
matters: recall falls at 0.60 on the semantic axis, 0.66 on the final one.
floor recall@5 precision@5 emptyOnAnswerable nonEmptyOnRefusal
0.00 0.991 0.221 0% 100%
0.62 0.991 0.524 0% 60%
0.66 0.991 0.695 0% 60%
0.67 0.972 0.704 1% 60%
Default 0.62, not the last-lossless 0.66: the refusal win is complete at
0.62 and 28 documents is a thin basis for spending margin. 0.60 is the
honest floor. Three of five refusal goldens are near-corpus by
construction, where declining is the generation tier's job.
Both eval tiers pin the floor to the constant and ignore the setting, for
the same reason the harness never sets skipRetrieval.
4 metrics improved against baseline, no regressions.
Chat was run through Ollama's OpenAI-compatible /v1 endpoint, which has no
context-size field, so the num_ctx we sent was silently discarded and the
entire NUM_CTX_LADDER was dead code. Every conversation ran at the backend
default and was truncated from the middle (where history and
retrieved context actually sit). Measured with identical ~7k-token prompts:
2,060 tokens processed against 7,049 on the native endpoint, and with the
needed fact in the discarded middle the old path hallucinated a plausible
wrong number.
Transport: add a native /api/chat path using the already-installed ollama
package, sending a real options object (num_ctx, num_predict, temperature,
seed) plus keep_alive. The OpenAI path is still used for LM Studio and llama.cpp,
minus the ineffective num_ctx and with include_usage so both paths return
prompt token counts. /api/show is now read for trained context, parameter
size, quantization and any modelfile num_ctx. The <think> stream normalizer
moves to app/utils/think_stream.ts so both transports share it; it gains a
flush() that fixes a dangling partial tag being dropped at end of stream.
Token accounting: replace chars/N with a structural segmenter scaled by a
per-model factor learned by EWMA from the prompt_eval_count every response
already returned and threw away. End-to-end estimator error is 4.6% against
18-20% for a fixed divisor — qwen2.5 converges to k=1.26 where llama3 sits
at 1.00, so no single global constant serves both.
Context window: resolved once per model from trained context, the exact KV
cost derived from GGUF metadata (head_count_kv, not head_count — GQA makes
them differ 4x), and available VRAM, then memoized. Memoization is
load-bearing: asking for a different num_ctx makes Ollama unload and reload
the model, stalling the turn and discarding the KV cache. The floor never
overrides the model's trained ceiling — tinyllama trains to 2048, and a
blind 4096 floor pushes RoPE past anything the weights saw.
Prompt budget: planPrompt allocates explicitly instead of delegating
overflow to silent backend truncation — response reserve, then fixed blocks
(system and current query, never dropped), then RAG chunks whole and
best-first, then history in whole user+assistant turns, evicted newest-first
in blocks with an elision marker. Block eviction keeps the prefix stable for
several turns rather than shifting on every one.
KV-cache ordering: the per-turn retrieved block moves behind history to the
tail, so the stable prefix stays byte-identical turn to turn. At turn 14 of
a long conversation, prefill is 231ms at the tail against 539ms at the front
(~1ms/turn growth against ~25ms). The win is conditional — with a large
retrieved block and short history the two measure the same — so both
orderings ship behind RAG_PLACEMENT. The per-turn query rewrite also moves
to the tasks model; on a single-slot Ollama it was evicting the chat model's
cached prefix on every single turn.
Also: OLLAMA_CONTEXT_LENGTH and OLLAMA_KV_CACHE_TYPE=q8_0 on the container
(q8_0 is treated as headroom, never as a budgeting assumption, since
unsupported architectures fall back to f16 silently), a Context Window
setting in Settings > Models, and budget fields on PipelineTrace.
Eval: golden suites (--suite) so the new long_context goldens cannot move
the aggregates the committed baselines were recorded against, plus budget
and estimator-error metrics through the report path. Retrieval is unchanged
and eval:compare reports no metric outside the tolerance band; the
long_context suite scores correctness 1.000 with historyElided 0.500 (the
two eviction cases, by design). Unit tests 337 -> 386 passing, same 7
pre-existing failures.
Ingest-side CHAR_TO_TOKEN_RATIO is deliberately untouched: changing it
re-chunks the corpus and invalidates retrieval baselines, which belongs in
its own measured change.
Allows users to optionally specify a model to use for ancillary tasks (chat suggestions, chat titles, etc.) instead of strictly using the last used model. This means aesthetic/non-critical work can be passed to a small, lightweight model instead of a heavy reasoning model used for chats.
If a model is not selected in the AI Assistant settings (`/settings/models`), the existing behavior of using the last used model is retained.
An install that never went online while visiting /maps never provisions
world.pmtiles (it is lazily extracted from the internet), leaving the map
blank/grey offline outside any downloaded regional extracts.
Detect the world basemap's presence on disk and surface it to the UI so:
- /maps shows a clear notice explaining the base map isn't downloaded and
how to fix it, instead of a silent grey screen.
- /settings/maps offers an explicit "Download Base Map" action (~15 MB) to
provision it deliberately while online.
Adds MapService.checkWorldBasemapExists() / provisionWorldBasemap(), a
POST /maps/setup-world-basemap endpoint, and threads worldBasemapExists
through the maps page props.
Refs #1030
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(AI): per-model thinking toggle with global default (off)
Stop forcing thinking on for every capable model. Adds a global default
(ai.autoThinking, ships OFF) and a per-model override in the chat window,
shown only for thinking-capable models and remembered client-side.
The /v1 (OpenAI-compat) endpoint ignores `think`; reasoning_effort is the
real lever. The controller resolves per-request preference -> global default
-> OFF (gated on capability), and the service maps that to reasoning_effort
('none' to suppress on a capable model, 'medium' for gpt-oss, unset to let a
capable model default thinking on). A new thinkingCapable flag keeps
non-Ollama backends from ever receiving reasoning_effort.
installed-models is enriched with a `thinking` flag (checkModelHasThinking,
now memoized) so the picker knows which models get the toggle.
Stacks on #1078 (reasoning-field read + client-disconnect abort).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(AI): thinking toggle as NOMAD Switch + info tooltip
Address review feedback on the per-model thinking control:
- Use the shared Switch component (matches AI Assistant settings) instead
of a raw checkbox.
- Label "Thinking:" with a colon to match the adjacent "Model:" label.
- Add an InfoTooltip explaining what thinking does and where the default
lives. Extend InfoTooltip with optional `position` ('top'|'bottom') and
`align` ('center'|'right') so it opens downward and expands leftward from
the right-edge header slot instead of being clipped/crushed against the
viewport edge. Defaults preserve existing (benchmark page) behavior.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(creator-packs): backend rail for gated creator content packs
Phase 1 of Creator Packs: the app-side download rail for per-creator
video ZIMs served from the private R2 bucket behind the entitlement
Worker. No UI yet (Phase 2).
- Thread optional requestHeaders through the download util + job so the
ZIM fetch can carry `Authorization: Bearer <APP_KEY>` on both the HEAD
probe and the GET stream (Worker streams the ZIM directly, gated).
- Add resourceMetadata.skip_embedding and gate the RAG/EmbedFileJob
branch on it so video ZIMs are never sent to the knowledge base.
- New creator_packs ManifestType + creatorPacksSpecSchema (display
metadata only, no url) + CollectionManifestService wiring and
getCreatorPacksWithStatus() status join.
- collections/creator-packs.json catalog with the two seed packs.
- CreatorPackService.installPack: ensure Kiwix (auto-install if absent),
resolve the stable Worker URL, dispatch the gated download with
skip_embedding. Typed result codes.
- CreatorPacksController + GET/POST /api/creator-packs routes.
- CREATOR_PACKS_APP_KEY / CREATOR_PACKS_WORKER_BASE env (release-injected
key; app degrades to a clear not_configured error when unset).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(creator-packs): Phase 2 UI — Content Explorer, settings page, wizard step
Adds the three user-facing surfaces for Creator Packs on top of the Phase 1
backend rail. All surfaces HIDE entirely on builds without the release-injected
key (configured=false) so a fork never shows a broken install button.
- Backend: CreatorPackService.isConfigured() + `configured` flag on the
GET /api/creator-packs response; /settings/creator-packs route + controller
action.
- api client + useCreatorPacks hook (shared cached source of configured/packs/
downloads for every surface) + CreatorPackCard (status badges, art fallback,
update pill) + CreatorPacksSection (install-on-click grid + confirm modal with
the license note), reused by the Content Explorer block and settings page.
- Content Explorer: Creator Packs section before Additional Content.
- Settings: new /settings/creator-packs manage page + conditional nav item.
- Easy Setup wizard: new Creator Packs step after Content. Replaces the fragile
hardcoded AI-skip math with a computed activeSteps list so the two optional
steps (Creator Packs + AI) navigate correctly in every on/off combination;
selection state, storage projection, review summary, and finish-install wired.
Browser-tested on NOMAD3 across all surfaces (configured + fork paths), incl.
live install→download→cancel and AI-on/off wizard renumbering.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(creator-packs): use branded pack banners on the cards
Replace the generic colored-block cards with the branded 1060x175 banners we
build into each pack ZIM, shown on both the settings page and the Easy Setup
wizard step (and the Content Explorer block, via the shared card).
- Bundle the seed-pack banners as local webp under admin/public/creator-packs/
so they render offline (no external image dependency, matching the existing
/rogue-support-banner.webp convention).
- CreatorPackCard is now banner-forward: the banner is the hero, with a compact
status/metadata footer (videos · size, Installed/Downloading/Selected/Install,
update pill). Falls back to a simple header if an image is ever missing.
- Prefer an optional catalog `banner_url` (future remote creators) over the
bundled-by-id path; add banner_url to the CreatorPack type + validator.
- Widen the card grids to 1–2 columns since banners are wide.
Browser-tested on NOMAD3: banners render on the settings page and wizard step;
selection/installed/downloading states and storage projection still correct.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(creator-packs): Phase 3 — uninstall, license, release key-injection
Completes the Creator Packs feature.
- Uninstall: CreatorPackService.uninstallPack reuses ZimService.delete (removes
the ZIM, drops it from the Kiwix library, clears the InstalledResource) behind
DELETE /api/creator-packs/:id. A trash control on installed cards appears only
on the settings "manage" surface (allowUninstall) — never on the Content
Explorer block or the wizard — with a danger confirm modal.
- License: draft "Project NOMAD Creator Pack License v1.0" at
collections/creator-pack-license.md (marked DRAFT — pending legal review;
personal-use, no-redistribution, official-channel-only, creators retain
copyright). The install modal links to it ("View license").
- Release key-injection: Dockerfile ARG/ENV CREATOR_PACKS_APP_KEY in the runtime
stage (empty default → source/CI builds ship unconfigured and hide the UI) +
build-primary-image.yml passes it from the CREATOR_PACKS_APP_KEY CI secret.
NOTE FOR JAKE: add that repo secret = the entitlement Worker's key.
- Update flow was already functional (installed-with-update cards → "Update
pack"); verified.
Browser-tested on NOMAD3: uninstall round-trip (file + DB row + kiwix library
cleaned, then reinstalled), settings-only uninstall control, install-modal
license link, and the "Update available" pill via a temporary catalog bump.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(downloads): clean up header merge logic
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: jakeaturner <jturner@cosmistack.com>
Adds a self-update path for the NOMAD admin/core image that runs without a
human in the loop, gated by layered safety checks. Recreation remains the
sidecar's job; this only decides *whether* to update now and requests it.
An hourly job (AutoUpdateJob) evaluates a side-effect-free decision pipeline
(AutoUpdateService.evaluate):
- opt-in: disabled by default
- in a user-configured local-time window (handles midnight wrap)
- an eligible release: same major (major bumps stay manual), strictly newer,
past a configurable cool-off, GA only (no drafts/prereleases), strict semver
- pre-flight: sidecar present, no in-flight system update / downloads / app
installs, and sufficient host disk (estimated from the registry manifest)
When all pass, it drives SystemUpdateService.requestUpdate() with a vetted tag.
Failure backoff auto-disables after 3 genuine update-request failures; transient
release-lookup errors are skips (offline-first appliances are routinely without
connectivity). Re-enabling clears the backoff state.
Settings UI exposes the toggle, window, and cool-off, with live status (eligible
target, in/out of window, last result/error). A `node ace auto-update:dry-run`
command exercises the full pipeline — and a deterministic --scenarios suite the
pure decision logic — without ever triggering an update.
New KVStore keys under autoUpdate.* hold config + last-run state.
New defaults:
OLLAMA_NO_CLOUD=1 - "Ollama can run in local only mode by disabling
Ollama’s cloud features. By turning off Ollama’s cloud features, you
will lose the ability to use Ollama’s cloud models and web search."
https://ollama.com/blog/web-searchhttps://docs.ollama.com/faq#how-do-i-disable-ollama%E2%80%99s-cloud-features
example output:
```
ollama run minimax-m2.7:cloud
Error: ollama cloud is disabled: remote model details are unavailable
```
This setting can be safely disabled as you have to click on a link to
login to ollama cloud and theres no real way to do that in nomad outside
of looking at the nomad_ollama logs.
This one can be disabled in settings in case theres a model out there
that doesn't play nice. but that doesnt seem necessary so far.
OLLAMA_FLASH_ATTENTION=1 - "Flash Attention is a feature of most modern
models that can significantly reduce memory usage as the context size
grows. "
Tested with llama3.2:
```
docker logs nomad_ollama --tail 1000 2>&1 |grep --color -i flash_attn
llama_context: flash_attn = enabled
```
And with second_constantine/deepseek-coder-v2 with is based on
https://huggingface.co/lmstudio-community/DeepSeek-Coder-V2-Lite-Instruct-GGUF
which is a model that specifically calls out that you should disable
flash attention, but during testing it seems ollama can do this for you
automatically:
```
docker logs nomad_ollama --tail 1000 2>&1 |grep --color -i flash_attn
llama_context: flash_attn = disabled
```
Adds a new settings page with Ko-fi donation link, Rogue Support
banner, and community contribution options (GitHub, Discord).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add comprehensive benchmarking capability to measure server performance:
Backend:
- BenchmarkService with CPU, memory, disk, and AI benchmarks using sysbench
- Database migrations for benchmark_results and benchmark_settings tables
- REST API endpoints for running benchmarks and retrieving results
- CLI commands: benchmark:run, benchmark:results, benchmark:submit
- BullMQ job for async benchmark execution with SSE progress updates
- Synchronous mode option (?sync=true) for simpler local dev setup
Frontend:
- Benchmark settings page with circular gauges for scores
- NOMAD Score display with weighted composite calculation
- System Performance section (CPU, Memory, Disk Read/Write)
- AI Performance section (tokens/sec, time to first token)
- Hardware Information display
- Expandable Benchmark Details section
- Progress simulation during sync benchmark execution
Easy Setup Integration:
- Added System Benchmark to Additional Tools section
- Built-in capability pattern for non-Docker features
- Click-to-navigate behavior for built-in tools
Fixes:
- Docker log multiplexing issue (Tty: true) for proper output parsing
- Consolidated disk benchmarks into single container execution
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
