Loom — Data Storyteller
Find interesting data, turn it into a readable chart, and share it — in your browser, on your phone, or as a macOS app.
▶ Try it: loom.ibm.io — no install. Live earthquakes, flights, Wikipedia edits, crypto, and weather are one click away, or load your own CSVs.
The desktop app (Tauri + DuckDB) adds folder mounting, millions of rows, SQL over CSV/Parquet, and WebGPU scatter. Charts are Vega-Lite specs drawn on Canvas / WebGPU; export PNG, SVG, or a link that reopens the exact chart.
Contributors & AI agents: See DOCS.md for architecture and AGENTS.md for a concise repo map and conventions. Platform support: docs/CAPABILITY_MATRIX.md. Verification gates: docs/DEPLOYMENT_GATES.md.
What it does
On your phone (web)
- A home for your data — Open Data library for My data, Bookmarks, Combinations, and Discover. Keep uploads and useful public sources close without copying them. Search your library, inspect a dataset’s origin, or browse one public catalog at a time. Select two datasets to build a saved lookup combination: choose matching fields, check matched and unmatched rows, and name the result. Combinations read their sources again when opened; charts keep the result from that open. Data stays on-device, and workspace backups include bookmark metadata and recipes. Details and limits.
- Find your next story — ✦ What’s interesting right now considers all 59 chart types across live and open feeds, from distributions and relationships to maps, networks, multivariate glyphs, and 3D views. Picks balance observed signals with variety across topics, sources, and questions. Browse Compare & rank, Over time, Distributions, Relationships, Parts of a whole, Connections, Maps & globes, and Many dimensions. Cards explain what stands out with measured evidence; stronger picks receive more space. Related views stay together in a perspective selector, while different questions and statistics get their own cards. Discover also considers category shapes and meaningful emoji for weather, disaster types, transit modes, and day/night fire observations. Use Emoji & shapes to find them; previews, legends, and shared links preserve their encodings. Search and filter types, time windows, facets, comparisons, and transforms; switch between Flow, Tiles, Masonry, and List. Open the previewed recipe and its exact snapshot, inspect provenance, or refresh for new ideas. Data library → Discover also offers curated packs and demos.
- More of what you like — Tap ♡ on a feed, pack, or story (⋯) to see more of it first in Discover, Data & sources, and search; 🔖 saves it to a Saved group; Less like this and Hide push it down or out (an "N hidden · Show" chip brings it back). Data & sources has one search box plus Topic / Where / Freshness / Saved filters and For you sorting. Manage everything in Settings → Your sources; picks stay on your device and travel in workspace backups.
- Fresh daily, fast to reopen — Picks are saved on your device and reused on recent visits. A new local calendar day triggers a background refresh while Loom is open; existing picks stay usable during updates. Manual Refresh checks again anytime. Chart taps display the saved data immediately and reconnect the feed afterward. Smaller phone type, tucked-away filters, and columns that fit the available width put more charts on screen.
- Stay in the story — Each card shows when its snapshot was checked. While Loom is open and its tab is active, a rotating check refreshes the whole source catalog on each feed’s cadence, with a full discovery scan each day and checks when you return online or to the tab. New data ready counts changed sources and lets you bring in updates without moving the cards you’re reading. Unchanged polls update the check time quietly; recently changed data gets a highlight. Closed apps do not refresh. Related chart perspectives share one card, with readable titles and a view selector. Back to stories keeps your search, filters, layout, and scroll position; browsing preferences survive a reload in the same tab.
- Make it yours — Edit opens a half-height sheet so the chart stays visible while you change encoding or look; ✦ Swipe ideas deals more charts to Keep / Skip. Bars flip horizontal on tall screens so long labels stay readable; taps show tooltips and finger-drags scrub them.
- Share — the Share button renders the chart at Square / Portrait / Story / Wide sizes with feed-sized type, then opens the phone’s share sheet (or saves the image). Get link publishes a chart page that unfurls with a preview image in iMessage, Slack, X, etc. (kept 7 days). Copy chart link (and the address bar in Chart view,
#chart=…) reopens the exact chart — dataset, chart type, encodings, headline, look, and framing; live feeds, demo files, and open-data CSVs reopen on load, local files wait for the recipient to open a file with the same name.
Explorer
- Reliable combinations — Matching-only combinations count the result toward the one-million-cell preview limit. Saved recipe uploads stay protected during routine local storage cleanup, including recipes stored with only a file reference.
- Saved file recovery — Older upload references retain storage protection. Reopening skips damaged or unsupported saved payloads without deleting them, and late saves cannot restore a file just deleted through its older reference.
- Touch controls — On phones, reorder columns with ← / → in the Columns menu. Dive offers Keep only / Exclude for value taps, and cube axes can be swapped by tapping two chips.
- Mount a folder — Point at a directory of CSV/Parquet files; Loom scans and exposes them in the sidebar. Search files by name.
- Data table — Virtualized table with sort, column reorder (drag headers), show/hide columns, per-column text and range filters. Sparklines, value bars, heat tint, trend cues, and null % in headers. Date columns auto-formatted.
- Row selection — Checkboxes, keyboard (↑/↓ + Space). Export selected rows to CSV. Saved views store column visibility, order, and filters; Undo/Redo for table layout.
- Linked highlighting — Hover a row in the table to highlight the corresponding point on the chart (and vice versa via scatter hover).
- Column profiling — Right-click (or long-press) a column header for a quick profile: null %, unique count, min/max/median, distribution histogram, top values.
Chart
- Refine without losing your work — Encoding edits retain your headline, notes, reference lines, and channel inversions through Undo/Redo. Heatmaps and hexbins can color by row count or a numeric measure with sum, mean, min, or max. Missing measurements stay missing; signed bars and stacked areas keep gains and losses on opposite sides of zero. Grouped and stacked bars follow the selected direction, and explicit grid, label, opacity, and line-style choices take precedence over automatic styling.
- Experiment and keep your style — The chart panel has Undo/Redo for encoding and visual edits (⌘/Ctrl Z and ⌘/Ctrl Shift Z outside text fields). Rapid slider changes group into one edit. In Visual → My looks, save up to 20 named styles, reuse them across datasets, and include them in workspace backups. Visual sections collapse individually or together, and color palettes are searchable.
- Find a field quickly — Encoding fields support search by name or type and filters for numeric, text, and time columns. Use arrow keys, Home/End, and Enter to choose; incompatible fields are skipped. Menus fit the visible screen when editing on a phone.
- Chart view — Pick a file, get instant chart suggestions (bar, line, scatter, area, pie, heatmap, strip, box). Click a suggestion or use Suggest with AI (Ollama).
- Switch types without losing the plot — changing chart type keeps fields that suit the new type and swaps in a fitting column where they don't (a scatter's numeric axis won't become 500 bar categories; glyph charts pick a name column; line / area / bump prefer a time column). Long daily series roll up to weeks or months automatically.
- Readable by default — Axes tick at round numbers with gridlines on the same values; dates read as months and years. Category names are never chopped: bars turn sideways when names won’t fit, crowded axes angle or thin their labels. Hover any bar, slice, bin, or cell for its actual value. One color unless Color encodes a field — then a legend appears in the emptiest corner. Heatmap, hexbin, and map ramps work in light and dark themes.
- Encode your way — X, Y, Color, Size, Row; plus Glow by, Outline by, Opacity by, Shape by category, and Emoji from column. Shape and emoji encodings work on scatterplots and point/bubble/globe maps. Bar stacking: grouped, stacked, or 100% stacked. Scatter: connect points (trail), marginal distributions.
- Arrange the canvas — Click Arrange above a chart to move or resize its title, legend, plot and notes. Resize handles keep the canvas in bounds; social aspect presets retain their shape. Arrows nudge, Shift makes larger steps, Ctrl/⌘ + arrows resize, and Escape cancels a drag. Each gesture is one Undo step. PNG exports preserve the arrangement.
Dive (Scuba-style slice and dice)
- Ask questions by clicking — pick a time window (last 15 min → last year, ending at the newest row, or Custom with absolute or relative bounds like
-3 hours / yesterday), filters, group by (nested), and metrics (count, count distinct, sum, avg, min, max, p5 … p99.9; min / max / avg also work on time columns as first / last seen). Results update instantly.
- Filters — multi-value chips with suggestions from the data (= ORs them, ≠ excludes them all), contains, LIKE, and regex.
- Derived columns — define a column with a DuckDB-style expression (
hour(ts), lower(country), CASE WHEN … END, regexp_extract(…)) and filter / group / aggregate on it. Evaluated in the browser; the same text goes into the SQL preview.
- Three views of one query — Time series (auto, fine, or fixed buckets from 1 s to 30 days; one line per top group; one panel per metric; empty buckets as 0, gaps, or connected), Table (ranked groups with totals, hits + share, highest- or lowest-first), Samples (newest matching raw rows, pick columns, sort by any header).
- Compare with the previous period, 1 day, 1 week, or 4 weeks earlier — dashed ghost lines and green/red % deltas.
- Drill in — click any value to filter to it (and stop grouping by it); ⌥-click to exclude it. On the chart: Break down by / Drill up, hover highlights the nearest line, click pins the crosshair, drag to zoom (double-click to zoom out).
- History — each change is a browser history step, so Back / Forward walk through your queries.
- Live — on Wikipedia / USGS / NWS and other feeds the dive auto-refreshes.
- Share — the whole query lives in the URL (
#dive=…); links to live feeds and demo files reopen the dataset on load. SQL shows the equivalent DuckDB query (open it in Query on desktop). Open as cube sends the filtered rows to the Data cube.
- Works on web and desktop: the engine runs in the browser over the loaded rows (desktop pulls a 30k-row DuckDB sample first; web streams use their 8k-row buffer; uploaded CSVs on web use the 500-row sample).
Monitor
- Watch any live feed: pick a number (latest, average, max, 95th percentile, count…), filter rows (city = "New York"), and choose when it matters: a limit, a healthy range, a sudden % change, unusual for this hour or weekday, or a feed that went quiet.
- Tied to decisions: every monitor says what you'll do when it fires. Each alert asks what actually happened (acted, expected, done on purpose, false alarm), and the card tells you when a monitor is mostly noise.
- Expected is not wrong: add maintenance windows, planned events, or recurring quiet hours. Mark an alert "expected" and Loom stops alerting at that weekday and hour.
- Scorecards: group related metrics (temperature + humidity + air quality) and see the worst status at a glance.
- Schedules: every N minutes, daily, weekly, or cron (
0 7 * * 1-5). Confirm over several checks to skip one-off blips, set reminder cooldowns, and set quiet hours.
- Notifications: in-app banners plus system notifications (desktop, iPhone/iPad, browser). Always on monitors are checked by Loom's server every 5 minutes even with the app closed, and alert your ntfy topic, Slack, Discord, or a webhook.
Query
- SQL editor — Run DuckDB SQL against
loom_active. Schema browser (Tables + Columns) and click-to-insert. Validation (parentheses, SELECT/WITH) before run.
- Results — Paginated grid, copy cell/row, export CSV. Query history and snippets (save/load named SQL). Snapshot current result and Diff vs a snapshot (row count delta).
- NL-to-SQL — Plain-language input (e.g. “show me sales by region”); scaffold query with schema context (full generation via Ollama when available).
App
- Theme-specific linework — Soft themes use rounded line ends and corners; Paper uses flat ends and beveled joins; Brutal, Tank, and NES use square ends and sharp joins. Visual → Marks lets you override Line endings and Line corners. Smooth curve remains a separate choice.
- Pair type with your theme — Settings → Follow theme coordinates typeface and chrome as you change themes. Manual choices keep your own combination. Source Sans 3, Source Serif 4, and IBM Plex Mono join the app and chart font menus; device fonts remain the default, with optional downloads. Paper, Glass, Frost, and geometric themes also have distinct surface shadows and icon weights.
- Themes that carry through — Discover’s palette button cycles through all ten themes. Automatic category colors follow the theme in previews and charts; theme-specific marks, line weights, and bar corners provide a starting point for Visual edits. Explicit palettes and accessibility color choices remain available.
- Theming — Dark, light, high-contrast, colorblind; font scale; reduced motion. Tokens in
globals.css.
- Onboarding — First-run modal: add data, then explore.
- Data & sources — Data.gov / data.gov.uk catalog search, sort, larger result sets (up to 200), and a quick page filter (Tauri); save CSV to folder; Wikipedia live stream (SSE →
wiki_stream); 60 live and poll feeds in six groups — Earth & climate (USGS quakes, NASA natural events, GDACS disaster alerts, weather and air quality for 12 cities, UK grid carbon, global temperature since 1880, ocean buoys, VIIRS active fires, USGS river gauges), Cities & transport (live aircraft, Citi Bike docks, bike share in Chicago / DC / SF, Boston MBTA vehicles, NYC 311, US airport delays, NYC restaurant inspections and parking tickets, Chicago crime reports), Space (ISS, upcoming launches, SpaceX history, NOAA space weather, aurora forecast, asteroid flybys, Starlink constellation, fireballs, visible satellites), Web & markets (Hacker News, Lobsters, Wikipedia most-read, Steam top games, crypto, Bitcoin blocks, 90 days of FX, internet outages, live BGP routing, Bluesky pulse, GitHub activity, new arXiv papers, trending books, Open Trivia, Gen‑1 Pokémon, internet radio), United States (NWS alerts, US national debt, every state year by year and the US economy month by month via USAFacts, FEMA, FDA and vehicle recalls, lobbying filings, federal agency budgets, Powerball draws), (COVID, countries, World Bank, Nobel laureates, Met Museum highlights, Open Food Facts) — each opens with charts picked for it and has Query-view SQL against virtual sources. The Data & sources browser has a shortlist and a full . load whole public CSVs: OWID CO₂, life expectancy, GDP, renewables, and population; the Keeling curve; every confirmed exoplanet; 35k power plants; US cities; Gapminder; every M7+ earthquake since 1900; Nobel laureates; FiveThirtyEight's candy ranking. searches every Socrata portal (NYC, Chicago, SF, Seattle, Austin, states, CDC…) by keyword, portal, and sort, and Explore pulls up to ~50k rows of the table straight into Chart. lists every weekly community dataset since 2018 (newest first, filterable by topic, file, or year) — tap a CSV/TSV file to chart it.
Screenshots
All captured from loom.ibm.io — each chart opens from a shareable #chart= / #dive= link.
| Live earthquakes on a map | Dive: Wikipedia edits by wiki |
|---|
| |
How these were captured (and how to refresh them): docs/screenshots/README.md.
Quick start
Prerequisites
- macOS 14+ (Sonoma) with Apple Silicon recommended
- Rust 1.75+ (
rustup)
- Node.js 24.x (e.g.
nvm)
- Tauri CLI:
cargo install tauri-cli
- Docker (optional, for containerized web UI)
First time
make setup # install npm + cargo deps
make spool # generate sample data in .loom-data
make spin # launch Loom (Tauri + Next.js + optional Ollama)
Then Choose folder → pick .loom-data (or any folder with CSV/Parquet), select a file, and switch to Chart to see suggestions.
Command reference (the Loom)
Run make (or make help) to list all commands. Every target uses a weaving metaphor.
Develop
| Command | Description |
|---|
make spin | Full dev: Tauri + Next.js hot reload; starts Ollama in background if available |
make thread | Frontend-only dev (no Rust) — good for UI work and web-only testing |
npx wrangler dev --port 8787 + LOOM_DEV_API=http://localhost:8787 npm run dev | Web dev with the Worker’s /api/* + /s/* (live feeds, catalog search, share links) proxied locally |
make warp | Rust backend type-check only (cargo check) |
make setup | First-time: install npm deps and fetch Rust deps |
Build
| Command | Description |
|---|
make weave | Production build → .app bundle |
make weave-web | Static web export only (output in ./out) |
make weave-rust | Rust release binary only |
Data
| Command | Description |
|---|
make spool | Generate sample datasets (10K–1M rows) in .loom-data |
make spool-small | Small test set (1K rows) |
make spool-mega | Stress test (5M rows, ~500MB) |
Docker
| Command | Description |
|---|
make shuttle | Build and run web UI in Docker |
make shuttle-build | Build Docker image only |
make shuttle-down | Stop containers |
make shuttle-shell | Shell into running container |
Cloud (Cloudflare)
| Command | Description |
|---|
make loft | Build static export and deploy to Cloudflare Workers |
make loft-mcp | Deploy Loom MCP / ChatGPT plugin Worker |
make loft-dry | Same build, dry-run deploy (no upload) |
make thread-mcp | Local MCP Worker on :8787 |
Hosted at loom.ibm.io — the browser web UI. Live feeds (USGS, Wikipedia, NWS, weather, flights, crypto, Hacker News…), catalog discovery (Data.gov / UK), and published story links run through the Cloudflare Worker (workers/catalog-proxy.ts); CSV Load is in-memory and size-capped. Full DuckDB folder workflows need the desktop app.
npx wrangler login # once
make loft # → https://loom.ibm.io/
make loft-mcp # → https://loom-mcp.mhsenkow.workers.dev/mcp
MCP / ChatGPT plugin
Remote MCP (same pattern as wordcount): chart profiling + recommendations for ChatGPT Developer Mode, Claude connectors, Cursor via mcp-remote.
make loft-mcp
Custom domain (e.g. loom.ibm.io): Cloudflare dashboard → Workers & Pages → loom-storyteller → Domains → add the hostname (or set routes in wrangler.jsonc).
Live streams (web): Connect works in the browser — poll sources via /api/source/*, Wikipedia via EventSource. Buffers are in-memory (not DuckDB); use Explore after Connect.
Feedback notes: Bottom-left note FAB → GitHub issue. One-time: create a fine-grained PAT with Issues: Read and write on mhsenkow/Loom_story_teller, then:
npx wrangler secret put GITHUB_TOKEN
Without the secret, notes still open the GitHub “new issue” form as a fallback.
Quality
| Command | Description |
|---|
make test | Run unit tests (Vitest) |
make check | Run all checks (TypeScript + Rust) |
make check-ts | TypeScript type-check + lint |
make check-rust | Rust check + clippy |
make fmt | Format code (Prettier + rustfmt) |
Cleanup
| Command | Description |
|---|
make unspool | Remove generated .loom-data |
make unravel | Deep clean (node_modules, .next, out, targets, data) |
make tidy | Light clean (build caches only) |
All of these are also exposed as npm scripts: npm run spin, npm run weave, npm run spool, etc.
Ollama (optional) — AI chart suggestions
To use Suggest with AI in the Chart view:
- Run Ollama and pull a model:
ollama serve
ollama pull llama3.2
- Optional env (
.env.local or shell):
NEXT_PUBLIC_OLLAMA_URL — default http://localhost:11434NEXT_PUBLIC_OLLAMA_MODEL — default: first available or llama3.2
- If the app can’t reach Ollama (e.g. CORS), allow origins:
OLLAMA_ORIGINS="*" ollama serve
Or set OLLAMA_ORIGINS to your dev URL (e.g. http://localhost:1420).
Hover Why? on a suggestion to see the reason (heuristic or AI).
Architecture
┌──────────────────────────────────────────────────────────────────┐
│ Tauri Shell (Rust) │
│ ┌──────────────┐ ┌──────────────┐ ┌─────────────────────────┐ │
│ │ tauri-fs │ │ DuckDB │ │ reqwest (Data.gov, │ │
│ │ (folder I/O) │ │ (analytics) │ │ save CSV to folder) │ │
│ └──────┬───────┘ └──────┬───────┘ └───────────┬─────────────┘ │
│ │ IPC (invoke) │ │ │
├─────────┼─────────────────┼────────────────────────┼──────────────┤
│ Frontend (Next.js + TypeScript) │
│ ┌──────┴───────┐ ┌──────┴───────┐ ┌─────────────┴───────────┐ │
│ │ Zustand │ │ Vega-Lite │ │ WebGPU / Canvas 2D │ │
│ │ (state) │ │ (spec gen) │ │ (scatter + bar/line/…) │ │
│ └──────────────┘ └──────────────┘ └──────────────────────────┘ │
└──────────────────────────────────────────────────────────────────┘
- Vega-Lite (brain) — Declares what to draw as portable JSON specs; used for export and for non-WebGPU marks.
- WebGPU / Canvas (muscle) — Renders scatter at scale; other mark types use Canvas 2D or Vega headless where appropriate.
Project structure
Loom_story_teller/
├── src-tauri/ # Rust backend
│ ├── Cargo.toml # DuckDB, Tauri, reqwest, etc.
│ ├── tauri.conf.json # Window, plugins, CSP
│ ├── capabilities/ # Tauri v2 permissions
│ └── src/
│ ├── lib.rs # Plugin init, command registration, DB + stream + sources state
│ ├── main.rs # Binary entry
│ ├── db.rs # DuckDB: scan, query, column stats
│ ├── stream.rs # Wikipedia SSE → wiki_stream
│ ├── sources.rs # Poll sources → DuckDB tables (schema from sourceRegistry.ts)
│ └── commands.rs # All #[tauri::command] handlers (see tauri.ts)
│
├── src/ # Next.js frontend
│ ├── app/
│ │ ├── layout.tsx # Root layout, fonts, globals
│ │ └── page.tsx # Three-panel layout (Sidebar | Main | DetailPanel)
│ ├── components/
│ │ ├── Sidebar.tsx # Files, Data & sources, live stream + poll source cards
│ │ ├── TopBar.tsx # View tabs (Explorer / Chart / Query)
│ │ ├── DetailPanel.tsx # Right panel: Stats, Chart (encoding + Visual), Export, Smart
│ │ ├── ChartView.tsx # Chart canvas, suggestions, Smart overlays, title edit, export
│ │ ├── ChartCard.tsx # Thumbnail + “Try” for suggestions
│ │ ├── ExplorerView.tsx # Full-width data table
│ │ ├── QueryView.tsx # SQL editor + results
│ │ └── PreviewFooter.tsx # Collapsible preview + Schema (drag tokens)
│ ├── lib/
│ │ ├── store.ts # Zustand state
│ │ ├── tauri.ts # Typed IPC bridge (invoke wrappers)
│ │ ├── vega.ts # Vega-Lite spec builders
│ │ ├── webgpu.ts # WebGPU pipeline (scatter)
│ │ ├── dive.ts # Dive engine: filters, group by, metrics/percentiles, buckets, compare, SQL, links
│ │ ├── diveSource.ts # Dive row loading (DuckDB sample on desktop, JS buffers on web)
│ │ ├── diveExpr.ts # Dive derived columns: DuckDB-style expression parser + evaluator
│ │ ├── dataCube.ts # Data cube binning, aggregation, camera, Canvas renderer, hover pick
│ │ ├── webgpuCube.ts # WebGPU instanced-voxel renderer for the data cube (cube.wgsl)
│ │ ├── recommendations.ts # Heuristics + recommendStreamStory / recommendSourceStory + SQL snippets
│ │ ├── ollama.ts # Ollama API for AI suggestions
│ │ ├── mock-data.ts # Browser fallbacks when not in Tauri
│ │ ├── format.ts # Number/byte formatting
│ │ ├── chartPalettes.ts # Chart color palettes
│ │ └── smartAnalytics.ts # Anomaly, forecast, trend, reference lines, clustering
│ ├── shaders/
│ │ └── scatter.wgsl # Compute + vertex + fragment
│ └── styles/
│ └── globals.css # Design tokens, theme
│
├── scripts/
│ └── generate_data.py # Sample data (scatter, sales, timeseries)
├── docs/
│ └── screenshots/ # Screenshots for README
├── Makefile # Command Loom (run `make` for help)
├── Dockerfile # Web UI container
├── docker-compose.yml
├── mcp/ # Remote MCP + ChatGPT plugin Worker
│ ├── src/index.ts # Tools: profile / recommend / build_chart
│ └── SUBMISSION.md # OpenAI plugin directory checklist
├── wrangler.jsonc # Cloudflare Workers static web UI
├── next.config.mjs # Static export for Tauri
├── tailwind.config.ts # Token-linked theme
└── package.json
See DOCS.md for a deeper codebase map and conventions.
Design system
Token-based in src/styles/globals.css, catalog in src/lib/lookSystem.ts (aligned with ibm.io wordcount / portfolio).
Themes (data-theme): light · dark · contrast · paper · glass · frost · brutal · loom · tank · nes
Chrome (data-ui): braun · monocle · bauhaus · noyes · ikea · military · terminal · nyt
Type: Libre Baskerville, Lora, IBM Plex, Inter, Geist, JetBrains Mono, Fira Code · faces auto/web/local
--loom-* aliases map to semantic colors so Tailwind (bg-loom-surface, …) keeps working. Chart → Visual presets replace look bundles (not merge). Colorblind chart series is a separate a11y toggle.
| Token | Purpose |
|---|
--loom-bg / --color-bg | Page background |
--loom-surface | Cards, panels |
--loom-elevated | Hover, inputs |
--loom-border | Borders |
--loom-text / --loom-muted | Text |
--loom-accent | Accent |
--chart-1 … --chart-8 | Series colors (per theme) |
--shadow-1 / --shadow-2 | Theme-aware shadows |
Component classes: .loom-panel, .loom-card, .loom-btn-primary, .loom-btn-ghost, .loom-input, .loom-badge.
IPC command reference
Frontend calls go through src/lib/tauri.ts; do not use raw invoke().
| Command | Args | Returns |
|---|
scan_folder | { folderPath: string } | FileEntry[] |
query_file | { filePath, sql, limit? } | QueryResult |
get_column_stats | { filePath: string } | ColumnInfo[] |
get_sample_rows | { filePath, limit? } | QueryResult |
inspect_file | { filePath, limit? } | InspectResult (stats + sample) |
save_csv_to_folder | { folder_path, url, filename } | string (saved path) |
fetch_data_gov_recent_csv, fetch_uk_data_recent_csv | { rows?, query?, sort? } | DataGovDataset[] (CKAN CSV search) |
stream_* | (see tauri.ts) | Wikipedia SSE ingest → wiki_stream |
source_* | kind: usgs | meteo | nws | world_bank | Poll-based APIs → per-source tables |
Full signatures live in src/lib/tauri.ts and Rust in src-tauri/src/commands.rs.
WebGPU pipeline (scatter)
WebGPU is used for scatter only when the mark is circle, there is no outline/jitter/glow, emoji, category-shape, data-driven glow/outline/opacity encoding, or Smart overlay (anomaly, trend, forecast, reference lines, clustering). When any of those are active, scatter uses Canvas 2D so overlays and encodings render correctly.
CPU: Float32Array (x, y, category, size_norm)
→ upload to GPU storage buffer
→ compute_positions (workgroups) × size_scale
→ screen-space coords + color
→ vertex_main (instanced quads)
→ fragment_main (circle + soft edge)
→ framebuffer
Shaders: src/shaders/scatter.wgsl. Palette aligns with --chart-* in CSS. Size scale (0.5–2×) multiplies size-encoded point radius in the shader.
Milestones
- M1 (Core) — Folder → DuckDB → WebGPU scatter. Target: 1M+ points at 60fps.
- M2 (Skin) — WebGPU texture → MLX sidecar for AI-styled charts (Apple Silicon).
- M3 (Share) — Bundle spec + assets for governed story sharing.
Build notes
- vega-canvas warning — Next.js may report
Module not found: Can't resolve 'canvas' from vega-canvas. This is an optional native dependency used by Vega in Node; the browser build works without it. You can ignore the warning or add canvas as an optional dependency if you run Vega in Node.
Key decisions
- Static export — Tauri expects a static frontend;
next.config.mjs uses output: "export". No server-side API routes at runtime.
- In-process DB — DuckDB runs inside the Rust process; no separate database server.
- Vega-Lite as spec — Charts are declarative JSON, so they’re auditable and LLM-friendly; used for export (SVG) and for non-WebGPU marks.
- WebGPU for scatter — High-density scatter uses compute shaders; other marks use Canvas 2D or Vega as needed.
- Zustand — Single store for UI and cached results; avoids Context re-render chains.
- Data.gov in Rust — Data.gov “recent CSV” is fetched by a Tauri command (reqwest) so it works without a Next.js API route.
Plot backgrounds use --chart-bg, mirrored by THEME_UI.bg for Canvas and exports and guarded by theme inheritance tests. This separates plotting surfaces from workspace colors (especially NES). Dialogs and sharing sheets inherit the selected corner geometry.
Automatic chart palettes resolve scale before accessibility: categorical uses Okabe–Ito and magnitude uses Cividis when colorblind defaults are enabled. Explicit palette choices are preserved; the Visual panel shows the resolved palette and theme-aware swatches. Device-font mode labels named faces as fallbacks.
Theme mark materials apply to Canvas bars and pies (including previews): flat, glaze, satin, ink, etched. Visual → Composition can override or restore inheritance. The fill path and hit geometry stay unchanged; density ramps remain flat. SVG material finishes remain flat approximations. Grids and title weights inherit theme defaults. Data cube wall/grid colors derive from theme ink; toolbar and Discover share a theme cycle with distinct icons.
Firefly fields
Firefly plots keep observations on fixed X/Y axes. Size encodes steady glow intensity; without Size, an encoded time field supplies recency. Without either, every observation has equal weight. Missing timestamps never become row-order recency. Colored cores and source-over halos remain readable on light and dark themes. Axis and grid styling applies to the live chart and Canvas captures. No perpetual twinkle runs, so static data does not imply change and idle charts spend no animation frames.
Chart → Visual → Color keeps Auto and the existing presets, with independent search and category/gradient browsing. Customize up to 16 swatches or gradient stops using native color wells and hex entry; fit the number of colors to a field’s distinct values, reverse the scale, or restore the preset. Swatches follow each chart’s category order (not named category assignments); extra groups repeat the palette. Counts describe the loaded field, before chart-specific Top N or grouping. Encoding → Group into chunks a crowded color or shape into ranges (numbers) or the largest categories plus Other, on every chart that uses those channels. Custom colors are included in Undo/Redo, saved looks, share links, PNG and SVG; SVG remains an approximation for chart types already marked as such. Axis lines, labels and mark outlines also accept custom colors.
