PATSTAT Explorer

mtc.berlin · EPO TIP

Patent analysis and visualization platform for the EPO PATSTAT database on Google BigQuery. Built for the EPO Technology and Innovation Programme (TIP) JupyterHub environment.

Quick Start (TIP)

The only file you need is PATSTAT_Explorer.ipynb.

1. Download it from this repo (or: File → Save Link As… on the raw link)
2. Upload it to your TIP JupyterHub home directory
3. Open it and run the single cell

That's it. The notebook clones this repository into ~/patstat_svelte, installs all dependencies (Node.js, MCP server, Python packages), builds the app, and starts three services. A clickable link to the PATSTAT Explorer appears when everything is ready.

First run (~1–2 min): clones this repo, installs dependencies, builds the app, starts all services. Subsequent runs (~10s): pulls latest changes via git pull, starts services. The build step is skipped if no source files changed.

To stop: run stop() in the notebook or restart the kernel.

What It Does

PATSTAT Explorer provides interactive views for patent data analysis — applicant trends, co-filing networks, citation flows, technology classifications, and an AI-powered query interface.

Views

AI Query Interface

The AI Query page (/query) supports two modes:

• Natural Language: Ask questions like "Top 10 applicants in solid-state battery patents since 2015". Claude discovers the PATSTAT schema via MCP tools (list_tables, get_table_schema, search_tables, get_table_samples, execute_query), builds SQL, executes it, and returns results. All intermediate steps are streamed live to the UI.
• SQL Mode: Write and execute BigQuery SQL directly.

BYOK (Bring Your Own Key): The Anthropic API key is entered once via the Settings dialog and stored locally in ~/.patent-navigator/config.json. The MCP server does not require an API key.

Patent Detail Sheet

Clicking any patent row (in Text Search, AI Query, or any table with appln_id) opens a side panel showing:

• Title, abstract, basic metadata (authority, filing date, kind, granted status)
• Applicants with country badges and links to applicant analysis
• CPC codes linking to Patent Classification Explorer
• Espacenet link

Name Consolidation

PATSTAT contains many name variants for the same organization (SIEMENS AG, SIEMENS LTD, SIEMENS CORP, etc.). The search page provides:

• Auto-suggest grouping by core name (strips legal suffixes like AG, GmbH, Inc, Ltd, SA, NV)
• Multi-select with custom parent label
• All consolidated names are carried through to every deep-dive view via URL parameters (names=A|||B|||C&label=Group)
• Collapsible display on all pages showing which names are included

Architecture

┌──────────────────────────────────────────────────────┐
│  Browser                                             │
│  SvelteKit 5 · Tailwind CSS 4 · shadcn-svelte       │
│  D3.js (Force Graph, Sankey) · LayerChart            │
└──────────────┬───────────────────────────────────────┘
               │ HTTP
┌──────────────▼───────────────────────────────────────┐
│  SvelteKit Node.js Server (Port 52080)               │
│  Proxies API requests, serves static assets          │
└──────────────┬───────────────────────────────────────┘
               │ HTTP
┌──────────────▼───────────────────────────────────────┐
│  Python Sidecar — FastAPI (Port 52081)               │
│  SQL validation, patent detail, co-occurrence,       │
│  settings, agentic NL-to-SQL loop (SSE streaming)    │
└──────┬───────────────────────────┬───────────────────┘
       │                           │
       │ BigQuery                  │ MCP (Streamable HTTP)
       ▼                           ▼
┌──────────────┐    ┌──────────────────────────────────┐
│  PATSTAT     │    │  mtc.berlin PATSTAT MCP          │
│  (BigQuery)  │    │  (Port 8082)                     │
│              │    │  Schema discovery & query tools   │
└──────────────┘    └──────────────────────────────────┘

Process Management

launch.py orchestrates the full startup:

1. MCP install — pip install --user from mtcberlin/mtc-patstat-mcp-lite (idempotent)
2. npm install — only if node_modules doesn't exist
3. npm run build — only if source files changed since last build
4. Start MCP server (Port 8082) — schema discovery and query execution tools
5. Start Sidecar (Port 52081) — FastAPI, connects to MCP and BigQuery
6. Start App (Port 52080) — SvelteKit production server

All processes are managed in-memory. stop() terminates everything cleanly.

Stateless Context via URL

There is no app-level state store. All analysis context is encoded in URL search parameters:

/applicant?names=SIEMENS%20AG|||SIEMENS%20LTD&label=Siemens&cpc=H01M&from=2015&to=2024

src/lib/context.ts provides parseContext(), contextToParams(), nameWhereClause(), and contextLink() for consistent handling across all views.

Tech Stack

API Endpoints

Project Structure

├── launch.py                    # Bootstrap & process management
├── sidecar.py                   # FastAPI backend (BigQuery + MCP + Claude)
├── package.json                 # Node.js dependencies & scripts
├── svelte.config.js             # SvelteKit configuration (node adapter)
├── vite.config.ts               # Vite + Vitest configuration
├── src/
│   ├── app.css                  # Tailwind theme (MTC branding, oklch tokens)
│   ├── lib/
│   │   ├── context.ts           # URL-based analysis context
│   │   ├── patstat.ts           # Sidecar API client
│   │   ├── csv.ts               # Client-side CSV export
│   │   ├── utils.ts             # cn() utility
│   │   └── components/
│   │       ├── consolidated-names.svelte
│   │       ├── patent-detail-sheet.svelte
│   │       └── ui/              # shadcn-svelte components
│   │           ├── badge/       ├── button/      ├── card/
│   │           ├── checkbox/    ├── collapsible/  ├── dialog/
│   │           ├── input/       ├── separator/    ├── sheet/
│   │           ├── table/       ├── tabs/         └── textarea/
│   └── routes/
│       ├── +layout.svelte       # Navigation header
│       ├── +page.svelte         # Home
│       ├── search/              # Applicant search + consolidation
│       ├── applicant/           # Applicant analysis (trends, jurisdictions, CPC)
│       ├── text-search/         # Full-text patent search
│       ├── technology/          # CPC code browser
│       ├── countries/           # Country comparison
│       ├── network/             # Co-applicant force graph
│       ├── citations/           # Citation Sankey diagram
│       ├── co-occurrence/       # CPC heatmap
│       ├── query/               # AI Query Interface
│       └── api/                 # Server-side API routes
│           ├── health/
│           ├── query/
│           ├── patent/[appln_id]/
│           ├── nl-to-sql/
│           ├── settings/
│           └── co-occurrence/
├── tests/                       # Python tests (pytest)
│   ├── test_sidecar.py          # Health, query, settings, patent detail, NL-to-SQL
│   ├── test_mcp_agent.py        # Tool conversion, agentic loop, max iterations
│   └── test_sql_extract.py      # SQL extraction from markdown/text
└── static/
    ├── logo.svg
    ├── logo.png
    └── robots.txt

Testing

# Frontend (Vitest — 59 tests)
npm test

# Backend (pytest — 40 tests)
python -m pytest tests/

# Both
npm test && python -m pytest tests/

Tests cover: URL context parsing, CSV export, API client, SQL validation, patent detail endpoint, settings CRUD, NL-to-SQL SSE streaming, MCP tool conversion, agentic loop behavior, SQL extraction from markdown/text, max iterations safeguard.

Development

npm run dev          # Vite dev server with HMR
npm run check        # TypeScript/Svelte type checking
npm run test:watch   # Vitest in watch mode

The sidecar must be running separately for the app to work:

python sidecar.py                    # Port 52081
patstat-mcp-lite --http --port 8082  # MCP server (optional, for AI Query)

Deployment

The app is designed for EPO TIP JupyterHub. It uses relative asset paths (paths.relative: true in svelte.config.js) to work behind the JupyterHub proxy at any base path.

Production builds are generated by npm run build into the build/ directory and served with node build/.

License

Internal — mtc.berlin · EPO TIP