v0.7.27 — bundled AI hotfix

v0.7.27 — bundled AI hotfix (chatml + latency + banner copy)

Three bugs from the first real use of the bundled-Qwen image in
v0.7.26:

1. Summarize / explain / ask AI ran forever

chat_format="qwen" isn't stable across llama-cpp-python versions —
on 0.2.90+ it's either unregistered or aliased to a template that
doesn't match Qwen 2.5's actual ChatML format. Result: the model
never emits the EOS token, so generation runs all the way to
max_tokens at ~5-15 tok/sec on a shared CPU — 30-60 s per call,
felt like a hang.

Fixed by:

• chat_format="chatml" — Qwen 2.5's actual conversation template,
  stable in llama-cpp-python.
• Explicit stop tokens <|im_end|> + <|endoftext|> as belt-and-
  suspenders against a GGUF metadata quirk suppressing the EOS.
• Greedy decoding (temperature=0.0) — deterministic + slightly
  faster sampler.
• n_ctx halved (4096 → 2048) — prompt eval scales O(ctx²) and our
  prompts comfortably fit 2 k.
• Per-method max_tokens caps tightened across the board.

Method	Before	After
summarize_result	320	180
explain_sql	300	200
optimize_sql	420	260
generate_sql	400	240
infographic_brief	480	300

Worst-case wall clock on a 5 tok/sec shared CPU: now ~35-60 s per
call instead of ~60-100 s plus runaway.

2. "Failed to get API key" banner for bundled

The admin AI page banner literally said "check that the API key is
valid" when any non-mock provider fell back to mock. Wrong copy for
bundled, which has no API key concept.

Fixed by a new fallbackHint(provider) helper that returns provider-
specific guidance:

Provider	Hint
bundled	verify the bundled GGUF model exists at /app/models/ and that llama-cpp-python is installed
ollama	verify the Ollama server is reachable at OLLAMA_BASE_URL and the configured model is pulled
openai / anthropic	verify the API key is valid and the configured model is correct

The registry's WARNING log line was also updated to drop the
misleading "Check API key + bundled SDK install" tail and list the
actual first-line failure mode per provider.

3. Pydantic schema-shadow warnings

UserWarning: Field name "schema" in "ResultContext" / "InfographicContext" shadows an attribute in parent "BaseModel" fired on every server start.
Silenced at the model level with ConfigDict(protected_namespaces=()).

Upgrade

docker pull ghcr.io/sanniheruwala/rednotebook-ai:v0.7.27

HF Space: bump the FROM tag in the Space's Dockerfile from
:v0.7.26 to :v0.7.27.

v0.7.26 — fix the v0.7.25 image build

v0.7.26 — fix the v0.7.25 image build

v0.7.25 was tagged but never produced a Docker image — the build
failed in both amd64 and arm64 runtime stages with:

ERROR: Failed to build installable wheels for some pyproject.toml
based projects (llama-cpp-python)

Root cause

llama-cpp-python upstream doesn't publish a complete wheel matrix
to PyPI. They publish prebuilt CPU-only wheels to a dedicated index
at https://abetlen.github.io/llama-cpp-python/whl/cpu covering
every supported Python × linux_{x86_64,aarch64} combination.

pip in the python:3.12-slim runtime stage couldn't find a matching
manylinux wheel on PyPI and fell through to a source build,
which dies because slim images have no compiler / cmake / make.

The fix

One-line change in the runtime stage:

 COPY --from=python-build /wheels /wheels
-RUN pip install --no-cache-dir /wheels/*.whl \
+RUN pip install --no-cache-dir \
+        --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu \
+        /wheels/*.whl \
     && rm -rf /wheels

PyPI stays the primary index. The CPU wheel index is consulted
secondarily for packages PyPI doesn't fully cover — pip can only
pull llama-cpp-python from there because nothing else lives in
that index.

v0.7.25 status

The v0.7.25 GitHub Release page exists for documentation purposes,
but no :v0.7.25 Docker image was ever published. Use :v0.7.26
as the first release with the bundled-Qwen feature actually built.

Upgrade

docker pull ghcr.io/sanniheruwala/rednotebook-ai:v0.7.26

HF Space: bump the FROM tag in the Space's Dockerfile from
:v0.7.24 (the previous shipping release) directly to :v0.7.26.
Skip :v0.7.25 — it doesn't exist.

v0.7.25 — bundled local AI by default

v0.7.25 — bundled local AI by default

The published Docker image now ships with Qwen2.5-Coder-1.5B-Instruct
baked in as the default AI provider. Bare docker run gets you real
SQL generation, summaries, and infographic narratives — no API key,
no Ollama install, no setup.

This is the change that lets us legitimately call the tool AI-native.

What's bundled

Component	Detail
Model	Qwen2.5-Coder-1.5B-Instruct, Q4_K_M quant (~1 GB on disk)
License	Apache 2.0
Runtime	llama-cpp-python (CPU-only, ~50 MB wheel)
Image size impact	600 MB → ~1.6 GB
RAM at runtime	~1.5 GB resident while loaded
Speed	~30-50 tok/sec CPU → 3-5 s per SQL gen. Apple Silicon Metal / GPU = ~instant
Cold start	~2-3 s mmap at app boot, then every request is warm

Behaviour matrix

Path	Default AI_PROVIDER	Model file location
Docker image	bundled (baked into ENV)	/app/models/qwen2.5-coder-1.5b-instruct-q4_k_m.gguf
pip install rednotebook-ai	mock (no model in wheel)	User points QWEN_MODEL_PATH at their own GGUF
HF Space (after bumping FROM tag)	bundled	Inherits from image

Quality + tradeoff framing

A 1.5B coder model writes plausible SQL on the bundled sample
notebook and clearly-degraded SQL on real warehouses. The admin AI
page now leads with the bundled option and explains the tradeoff in
plain English so users know when to graduate to OpenAI / Anthropic /
a bigger local model via Ollama.

Graceful degradation

Two failure modes, both fall back to mock with a single WARNING log
line — never crash the AI surface:

1. llama_cpp not installed: the side-effect import in
   server/main.py catches it and bundled stays unregistered.
2. GGUF file missing on disk: BundledAIProvider.__init__ raises
   FileNotFoundError; the registry's existing try/except in
   get_provider catches it.

Three new tests cover both paths.

Upgrade

docker pull ghcr.io/sanniheruwala/rednotebook-ai:v0.7.25

HF Space refresh: bump the FROM tag in the Space's Dockerfile to
:v0.7.25. First HF rebuild will be slower than usual (~5-10 min
extra) because the model layer downloads 1 GB from HuggingFace. Every
subsequent app-only release reuses the cached model layer.

Knobs

• AI_PROVIDER=bundled|openai|anthropic|ollama|mock — pick provider.
• QWEN_MODEL_PATH=/path/to/your.gguf — override the bundled GGUF
  (drop in Qwen 7B, Phi-3.5, etc.).
• BUNDLED_AI_THREADS=4 — CPU thread count for inference. Defaults
  to cpu_count - 1.

v0.7.24 — chart hotfix

v0.7.24 — chart hotfix

Two regressions in v0.7.23 that real use surfaced within hours of
shipping.

Downloaded PNG was blank

The live ECharts in the full-size chart card runs in SVG renderer
mode for results under 3k points (the common analyst-shaped result).
In SVG mode, ECharts' getDataURL({type:'png'}) doesn't reliably
emit a real raster — depending on browser and chart shape it returns
either a mis-typed SVG-data-URL or a partly-rendered transparent
PNG. The "startsWith data:image/png" defensive check would sometimes
pass anyway, sending the malformed URL straight to the file system.

Now: PNG export always goes through a headless canvas re-render,
regardless of the live renderer. Plus four quality fixes:

• Animation disabled in the headless option so we don't capture
  the chart mid-grow (the second-most-common blank-export mode).
• One requestAnimationFrame wait after setOption so the
  renderer has produced its final pixels before we read them.
• White background instead of transparent — Slack / Docs /
  Notion render behind a light surface anyway, and transparent
  PNGs were being mis-perceived as empty in some viewers.
• DataZoom stripped from the still image so the pan slider
  doesn't pollute the exported chart.

Cost: ~150ms extra per PNG export. Worth it for always-correct
downloads.

Customize popover extended beyond the viewport with no scroll

max-h-[70vh] was meaningless on short windows — Radix sizes the
popover to its natural content height first, and 70vh wasn't tall
enough on screens under ~700px.

Now the popover is a flex column with:

• max-h-[var(--radix-popover-content-available-height)] —
  Radix's collision detector populates this CSS variable with the
  actual largest height that fits without going off-screen. Tighter
  than 70vh and dynamic.
• Pinned header (shrink-0) — the title and Reset stay visible
  while you scroll.
• Inner scrollable body (overflow-y-auto overscroll-contain).
• max-w-[calc(100vw-24px)] so the popover doesn't overflow
  the screen on tiny phones.
• collisionPadding={12} so it keeps breathing room from
  screen edges.

Upgrade

docker pull ghcr.io/sanniheruwala/rednotebook-ai:v0.7.24

HF Space refresh: bump the FROM tag in the Space's Dockerfile to
:v0.7.24.

v0.7.23 — Chart tab grew up

v0.7.23 — the Chart tab grew up

This release replaces the empty axis-picker that used to greet every
result with a complete report-grade chart card. It rolls in two
feature sets — the auto-recommended chart grid (PR #19) and the
Share + Customize + inline editing surface (PR #20) — because the
v0.7.22 version was transient and never tagged.

The new Chart tab, in one screen

┌─────────────────────────────────────────────────────────────────┐
│ 1. You land on a 2×2 grid of suggested charts (no clicks needed)│
│ 2. Pick one → it expands to full-size with report chrome        │
│ 3. Rename in place, tweak palette / format / layout in Customize│
│ 4. Share → download PNG / SVG / clipboard / CSV                 │
└─────────────────────────────────────────────────────────────────┘

Auto-recommended chart grid

A pure-TypeScript heuristic engine reads your result and ranks chart
candidates by shape:

Roles	Chart
time + numeric	line / area
time + numeric + low-card category	stacked area
low-card category + numeric	bar (top categories)
2-8 distinct categories + numeric	donut for composition
two numerics	scatter (coloured by a category if present)
single numeric	histogram
single-row + single numeric	KPI
boolean + numeric	true/false bar

Each tile shows a live ECharts thumbnail with a one-line "why this
one" caption. "Try another set" cycles through up to 8 ranked
candidates. "Custom" opens the manual axis picker for the cases
the heuristic missed.

Deterministic — same data, same suggestions. ~1ms to compute, no LLM,
no round trip.

Report-grade chart chrome

Every full-size chart now has:

• 1px gradient accent strip across the top in the active palette.
• HTML header above the canvas with an auto-generated title
  ({Y} by {X}, {Y} over time, Distribution of {Y}, etc.), a
  one-line description ("Sum of revenue · grouped by region"), and a
  row-count badge.
• HTML footer with the aggregation method, optional "truncated"
  warning, query duration in ms/s, and a quiet "Made with
  RedNotebook" mark.
• The internal ECharts title is removed so typography (tabular nums,
  kerning, weights) stays consistent with the rest of the app.

Share menu

Every chart has a Share button that opens a dropdown with four export
paths:

• Download as PNG — high-res raster (pixel ratio 2, 1280×720
  export size regardless of on-screen size), paste anywhere.
• Download as SVG — vector, scales without blur.
• Copy image to clipboard — straight into Slack / Docs / email.
• Download data as CSV — the numbers behind the chart, header
  row matches the result columns, CSV-escaped values for commas /
  quotes / newlines / dates.

Cross-renderer fallback: ECharts emits only the format that matches
its active renderer (canvas → PNG, SVG mode → SVG). The export utility
detects this and briefly mounts a hidden 1280×720 echarts instance in
the desired renderer to grab the right format. User never sees the
second render.

Customize popover

A new Customize button next to Share opens a popover with five
sections that all persist into config.options:

• Title & description — free-text override of the auto-generated.
• Look — 5 palette presets (Brand / Ocean / Forest / Sunset /
  Mono) with swatch previews + 3 heights (Compact / Standard / Tall).
• Numbers — Y-axis format: Auto / Number / $ / %.
• Show — toggles for legend / gridlines / data labels (bar) /
  smooth lines (line/area) / area fill (line). Chart-type-specific
  toggles hide themselves when irrelevant.

Inline title and subtitle editing

Click the title or subtitle in the header → it becomes an inline
input, autofocused and text-selected. Enter or blur commits; Escape
reverts. Most users want to rename, not open a dialog.

Read-only consumers (visualization-cell, published HTML) pass no
onChange and the field stays display-only.

Stable option memo

Title/subtitle changes used to trigger an ECharts canvas redraw +
re-animation per keystroke. The useMemo deps are now the structural
fields only (chart_type, x, y, color, aggregation,
options, filters) — HTML header changes no longer disturb the
canvas. Typing feels instant.

Notes

• The v0.7.22 version was transient and never tagged; this release
  covers everything that would have shipped under it.
• No new Python dependencies. One new frontend file
  (frontend/lib/chart-export.ts) and one new UI primitive wrapper
  (frontend/components/ui/popover.tsx over the already-installed
  Radix Popover).

Upgrade

docker pull ghcr.io/sanniheruwala/rednotebook-ai:v0.7.23

HF Space refresh: bump the FROM tag in the Space's Dockerfile from
v0.7.21 to v0.7.23.

v0.7.21

Install

Docker (any OS):

docker run -p 8000:8000 ghcr.io/sanniheruwala/rednotebook-ai:v0.7.21

Python (any OS): download the attached *.whl and pip install it.

See the README for full instructions.

What's Changed

• chore(deps): bump next from 14.2.35 to 15.5.18 in /frontend in the npm_and_yarn group across 1 directory by @dependabot[bot] in #7
• feat(onboarding+demo): quick-connect templates, tour, rich empty state, sample seed, DEMO_MODE (v0.7.21) by @sanniheruwala in #16

New Contributors

• @dependabot[bot] made their first contribution in #7

Full Changelog: v0.7.20...v0.7.21

v0.7.20

Install

Docker (any OS):

docker run -p 8000:8000 ghcr.io/sanniheruwala/rednotebook-ai:v0.7.20

Python (any OS): download the attached *.whl and pip install it.

See the README for full instructions.

What's Changed

• fix(dialog): drop duplicate close X + repair History dialog body height (v0.7.20) by @sanniheruwala in #13
• fix: complete the v0.7.20 patch — Cursor removal, persist on refresh, history redesign, publisher charts by @sanniheruwala in #15

Full Changelog: v0.7.19...v0.7.20

v0.7.19

Install

Docker (any OS):

docker run -p 8000:8000 ghcr.io/sanniheruwala/rednotebook-ai:v0.7.19

Python (any OS): download the attached *.whl and pip install it.

See the README for full instructions.

What's Changed

• feat(metadata+chart): virtual files catalog + HD chart renderer by @sanniheruwala in #11
• chore: bump to v0.7.19 by @sanniheruwala in #12

Full Changelog: v0.7.18...v0.7.19

v0.7.18

Install

Docker (any OS):

docker run -p 8000:8000 ghcr.io/sanniheruwala/rednotebook-ai:v0.7.18

Python (any OS): download the attached *.whl and pip install it.

See the README for full instructions.

What's Changed

• fix(docker): v0.7.18 hotfix — Alpine→Debian frontend stage + uploads/published dir chown by @sanniheruwala in #10

Full Changelog: v0.7.17...v0.7.18

v0.7.17

Install

Docker (any OS):

docker run -p 8000:8000 ghcr.io/sanniheruwala/rednotebook-ai:v0.7.17

Python (any OS): download the attached *.whl and pip install it.

See the README for full instructions.

What's Changed

• feat(sql-cell): dialect-aware SQL formatter button by @sanniheruwala in #8
• chore: bump to v0.7.17 by @sanniheruwala in #9

Full Changelog: v0.7.16...v0.7.17