Alternative: manual JSON config (if custom connectors aren't available)

\n\nAdd to `claude_desktop_config.json` ([Node.js](https://nodejs.org) required):\n\n```json\n{\n \"mcpServers\": {\n \"swiss-caselaw\": {\n \"command\": \"npx\",\n \"args\": [\"-y\", \"mcp-remote\", \"https://mcp.opencaselaw.ch\"]\n }\n }\n}\n```\n\nRestart Claude Desktop after saving.\n\n

\n\n**ChatGPT:**\n\n1. Open **Settings** → **Apps** → **Advanced settings** → enable **Developer mode**\n2. Click **Create app** → name it `Swiss Caselaw`, paste `https://mcp.opencaselaw.ch/sse`, auth: None\n3. In any chat: click **+** → **Developer mode** → select **Swiss Caselaw**\n\n> Available on Plus, Pro, Team, Enterprise, and Edu plans. Recommended with GPT-5.3 (GPT-5.4 does not currently support MCP tool invocation).\n\n**Gemini CLI:**\n\nAdd to `~/.gemini/settings.json`:\n\n```json\n{\n \"mcpServers\": {\n \"swiss-caselaw\": {\n \"url\": \"https://mcp.opencaselaw.ch\"\n }\n }\n}\n```\n\nRestart Gemini CLI after saving. No account plan required — Gemini CLI is free.\n\n**Google ADK (Agent Development Kit):**\n\nBuild a Gemini-powered agent with access to all tools:\n\n```python\nfrom google.adk.agents import LlmAgent\nfrom google.adk.tools import MCPToolset\nfrom mcp.client.sse import SseConnectionParams\n\nagent = LlmAgent(\n model=\"gemini-3.1-pro\",\n name=\"swiss_law_agent\",\n instruction=\"You are a Swiss legal research assistant.\",\n tools=[\n MCPToolset(\n connection_params=SseConnectionParams(\n url=\"https://mcp.opencaselaw.ch/sse\",\n ),\n ),\n ],\n)\n```\n\nSee the [full MCP setup guide](https://github.com/jonashertner/caselaw-repo-1/blob/main/docs/claude-desktop-setup.md) for Google Gen AI SDK examples and all other platforms.\n\n> See the **[full MCP setup guide](https://github.com/jonashertner/caselaw-repo-1/blob/main/docs/claude-desktop-setup.md)** for detailed instructions for all platforms.\n\n> The `update_database` and `check_update_status` tools are only available on the local server — the remote dataset is updated automatically every night.\n\n### Option B: Local server (offline access)\n\nRun the MCP server locally with your own copy of the database (~65 GB disk). This gives you offline access and full control over the data.\n\n#### Setup with Claude Code\n\n[Claude Code](https://docs.anthropic.com/en/docs/claude-code) is Anthropic's CLI for working with Claude in the terminal.\n\n**Step 1.** Clone this repository:\n\n```bash\ngit clone https://github.com/jonashertner/caselaw-repo-1.git\ncd caselaw-repo-1\n```\n\n**Step 2.** Create a virtual environment and install the MCP server dependencies:\n\n```bash\npython3 -m venv .venv\n\n# macOS / Linux\nsource .venv/bin/activate\npip install mcp pydantic huggingface-hub pyarrow fastapi\n\n# Windows (PowerShell)\n.venv\\Scripts\\Activate.ps1\npip install mcp pydantic huggingface-hub pyarrow\n```\n\n**Step 3.** Register the MCP server with Claude Code:\n\n```bash\n# macOS / Linux\nclaude mcp add swiss-caselaw -- /path/to/caselaw-repo-1/.venv/bin/python3 /path/to/caselaw-repo-1/mcp_server.py\n\n# Windows\nclaude mcp add swiss-caselaw -- C:\\path\\to\\caselaw-repo-1\\.venv\\Scripts\\python.exe C:\\path\\to\\caselaw-repo-1\\mcp_server.py\n```\n\nUse the full absolute path to the Python binary inside `.venv` so that the server always finds its dependencies, regardless of which directory you run Claude Code from.\n\n**Step 4.** Restart Claude Code and run your first search.\n\nOn first use, the server automatically:\n1. Downloads all Parquet files (~7 GB) from [HuggingFace](https://huggingface.co/datasets/voilaj/swiss-caselaw) to `~/.swiss-caselaw/parquet/`\n2. Builds a local SQLite FTS5 full-text search index at `~/.swiss-caselaw/decisions.db` (~58 GB)\n\nThis takes 30–60 minutes depending on your machine and connection. It only happens once — after that, searches run instantly against the local database.\n\n**Total disk usage:** ~65 GB in `~/.swiss-caselaw/` (macOS/Linux) or `%USERPROFILE%\\.swiss-caselaw\\` (Windows).\n\n#### Setup with Claude Desktop\n\nSee the **[Claude Desktop setup guide](https://github.com/jonashertner/caselaw-repo-1/blob/main/docs/claude-desktop-setup.md)** for step-by-step instructions (macOS + Windows).\n\n**Quick version** — add this to your `claude_desktop_config.json`:\n\n```json\n{\n \"mcpServers\": {\n \"swiss-caselaw\": {\n \"command\": \"/path/to/caselaw-repo-1/.venv/bin/python3\",\n \"args\": [\"/path/to/caselaw-repo-1/mcp_server.py\"]\n }\n }\n}\n```\n\nConfig file location: `~/Library/Application Support/Claude/claude_desktop_config.json` (macOS) or `%APPDATA%\\Claude\\claude_desktop_config.json` (Windows). On Windows, use `.venv\\\\Scripts\\\\python.exe` instead.\n\nAny MCP-compatible client works with the same `command` + `args` pattern.\n\n#### Keeping the data current\n\nThe dataset is updated daily. To get the latest decisions, ask Claude to run the `update_database` tool, or call it explicitly. This re-downloads the Parquet files from HuggingFace and rebuilds the local database.\n\n#### SQLite snapshot artifacts\n\nThe HuggingFace dataset may also publish a full compressed SQLite base snapshot for bootstrap tools under `artifacts/sqlite/snapshots/`. Its metadata lives in `artifacts/manifest.json` as `snapshot`; if that value is `null`, consumers should fall back to the Parquet rebuild path above.\n\nThe snapshot is intended for local MCP/server bootstrapping: download `artifacts/manifest.json`, fetch `snapshot.sqlite_zst.path`, verify `snapshot.sqlite_zst.sha256`, decompress to `decisions.db.tmp`, run a quick SQLite row/schema check, then atomically move it to `decisions.db`. Newer `artifacts/sqlite/deltas/*.sqlite.zst` entries can then be applied by tools that support delta updates.\n\nManual snapshot bootstrap:\n\n```bash\npip install zstandard\n\npython - <<'PY'\nimport hashlib\nimport json\nimport os\nimport sqlite3\nimport urllib.request\nfrom pathlib import Path\n\nimport zstandard as zstd\n\nbase = \"https://huggingface.co/datasets/voilaj/swiss-caselaw/resolve/main\"\ndata_dir = Path(os.environ.get(\"SWISS_CASELAW_DIR\", Path.home() / \".swiss-caselaw\")).expanduser()\ndata_dir.mkdir(parents=True, exist_ok=True)\n\nwith urllib.request.urlopen(f\"{base}/artifacts/manifest.json\") as r:\n manifest = json.load(r)\n\nsnapshot = manifest.get(\"snapshot\")\nif not snapshot:\n raise SystemExit(\"No SQLite snapshot is advertised; use update_database instead.\")\n\nmeta = snapshot[\"sqlite_zst\"]\ntmp_db = data_dir / \"decisions.db.tmp\"\nfinal_db = data_dir / \"decisions.db\"\n\nclass HashingReader:\n def __init__(self, raw, digest):\n self.raw = raw\n self.digest = digest\n\n def read(self, size=-1):\n data = self.raw.read(size)\n if data:\n self.digest.update(data)\n return data\n\nh = hashlib.sha256()\nwith urllib.request.urlopen(f\"{base}/{meta['path']}\") as src, tmp_db.open(\"wb\") as dst:\n zstd.ZstdDecompressor().copy_stream(HashingReader(src, h), dst)\nif h.hexdigest() != meta[\"sha256\"]:\n tmp_db.unlink(missing_ok=True)\n raise SystemExit(\"SHA-256 verification failed\")\n\ncon = sqlite3.connect(tmp_db)\ntry:\n rows = con.execute(\"SELECT COUNT(*) FROM decisions\").fetchone()[0]\n con.execute(\"SELECT decision_id FROM decisions LIMIT 1\").fetchone()\nfinally:\n con.close()\n\nexpected = snapshot.get(\"rows\")\nif expected is not None and rows != expected:\n raise SystemExit(f\"Row-count check failed: got {rows}, expected {expected}\")\n\ntmp_db.replace(final_db)\nprint(f\"Installed {final_db} with {rows:,} decisions\")\nPY\n```\n\n#### How the local database works\n\n```\n~/.swiss-caselaw/\n├── parquet/ # Downloaded Parquet files from HuggingFace (~7 GB)\n│ └── data/\n│ ├── bger.parquet\n│ ├── bvger.parquet\n│ └── ... # 100 files, one per court\n├── decisions.db # SQLite FTS5 search index (~58 GB)\n├── reference_graph.db # Case citations + statute links (~3.5 GB)\n└── statutes.db # Federal law articles from Fedlex (~610 MB)\n```\n\nAll data stays on your machine. No API calls are made during search — the MCP server queries the local SQLite database directly.\n\n**Database structure.** `decisions.db` is a single SQLite file with two tables:\n\n- **`decisions`** — the main table with one row per decision. Holds 24 columns, including the search-facing fields plus `json_data` (full record blob) and `canonical_key` for dedup-aware lookup. Indexed on `court`, `canton`, `decision_date`, `language`, `docket_number`, `chamber`, and `decision_type` for fast filtered queries.\n\n- **`decisions_fts`** — an FTS5 virtual table that mirrors 7 text columns from `decisions`: `court`, `canton`, `docket_number`, `language`, `title`, `regeste`, and `full_text`. FTS5 builds an inverted index over these columns, enabling sub-second full-text search across 1M+ decisions. The tokenizer is `unicode61 remove_diacritics 2`, which handles accented characters across German, French, Italian, and Romansh. Insert/update/delete triggers keep the FTS index in sync with the main table automatically.\n\n**Why ~58 GB.** The full text of 962K+ court decisions averages ~15 KB per decision. The FTS5 inverted index adds overhead for every unique token, its position, and the column it appears in. This is a known trade-off: FTS5 indexes over large text corpora are substantially larger than the source data, but they enable instant ranked search without external infrastructure.\n\n**Search pipeline.** When you search, the server:\n\n1. **Detects query intent** — docket number lookup (`6B_1234/2023`), explicit FTS syntax (`Mietrecht AND Kündigung`), or natural language (`decisions on tenant eviction`).\n\n2. **Runs multiple FTS5 query strategies** — For natural-language queries, the server generates several FTS query variants (AND, OR, phrase, field-focused on regeste/title, with multilingual term expansion) and executes them in sequence. Each strategy produces a ranked candidate set. For explicit syntax (AND/OR/NOT, quoted phrases, column filters), the raw query is tried first.\n\n3. **Fuses candidates via RRF** — Results from all strategies are merged using Reciprocal Rank Fusion: each candidate's score is the weighted sum of `1/(k + rank)` across all strategies that returned it. Decisions found by multiple strategies get a boost.\n\n4. **Reranks with signal scoring** — The top candidates are reranked using a composite score that combines:\n - **BM25 score** (from FTS5, with custom column weights: `full_text` 1.2, `regeste` 5.0, `title` 6.0 — headnotes and titles are weighted heavily over body text)\n - **Term coverage** in title (3.0x), regeste (2.2x), and snippet (0.8x)\n - **Phrase match** in title/regeste (1.8x)\n - **Docket match** — exact (6.0x) or partial (2.0x)\n - **Statute/citation graph signals** — if the query mentions an article (e.g., \"Art. 8 BV\"), decisions that cite that provision are boosted\n - **Court prior** — e.g., asylum queries boost BVGer results\n\n5. **Selects the best passage** — For each result, the server scans the full text for the most relevant passage and returns it as a snippet.\n\n### Available tools\n\nAvailable on both remote and local unless noted.\n\n| Tool | Description |\n|------|-------------|\n| `search_decisions` | Full-text search with filters (court, canton, language, date range, chamber) |\n| `get_decision` | Fetch a single decision by docket number or ID. Includes citation graph counts (cited by / cites). |\n| `list_courts` | List all courts with decision counts |\n| `get_statistics` | Aggregate stats by court, canton, or year |\n| `find_citations` | Show what a decision cites and what cites it, with confidence scores |\n| `find_appeal_chain` | Trace the appeal chain (Instanzenzug) — prior and subsequent instances |\n| `find_leading_cases` | Find the most-cited decisions for a topic or statute |\n| `analyze_legal_trend` | Year-by-year decision counts for a statute or topic |\n| `draft_mock_decision` | Build a research-only mock decision outline from facts, grounded in caselaw + statute references; asks clarification questions before conclusion |\n| `get_case_brief` | Structured brief for any case — regeste, key Erwägungen, cited statutes, and citation authority |\n| `get_doctrine` | Statute article or legal concept → ranked leading cases + doctrine timeline |\n| `generate_exam_question` | Legal topic → real BGE fact pattern with hidden analysis for Fallbearbeitung practice |\n| `get_law` | Look up a federal or cantonal Swiss law by SR number / abbreviation + optional canton. Returns full article text from the local mirror (Fedlex for federal; cantonal_laws.db for cantonal — direct portal scrape for 22 cantons, LexFind PDF extraction for the remaining 4) |\n| `search_laws` | Unified federal + cantonal FTS5 search with BM25 rank interleaving. Filter by `canton=` for a specific jurisdiction, or `jurisdiction='federal'` / `'cantonal'` |\n| `get_commentary` | Scholarly commentary for a specific federal law article from OnlineKommentar.ch (CC-BY-4.0) |\n| `search_commentaries` | Full-text search across 362 legal commentaries from OnlineKommentar.ch across 19 federal laws |\n| `search_legislation` | LexFind-backed discovery search across federal + cantonal legislation. `fetch_top_n_texts=N` enriches the top N results with full article text in a single call |\n| `get_legislation` | Full metadata + article text for any federal or cantonal law. Local-first (statutes.db → cantonal_laws.db → LexFind API fallback) |\n| `browse_legislation_changes` | Recent-changes feed per canton or federal level (live LexFind API, no local mirror) |\n| `update_database` | Re-download latest Parquet files from HuggingFace and rebuild the local database *(local only)* |\n| `check_update_status` | Check progress of a running database update *(local only)* |\n\n### Citation-integrity toolkit (anti-hallucination)\n\nOpenCaseLaw treats Swiss legal citations as a closed-corpus problem: every reference an LLM writes either resolves to a real entry in the ~1.05M-record corpus + Fedlex statute mirror, or it does not. The MCP server ships a four-tool toolkit that makes this contract enforceable end-to-end.\n\n| Tool | Purpose |\n|------|---------|\n| `cite` | Build the canonical citation_string + canonical_url for a Swiss reference. Returns `exists=false` plus close_matches when the reference is fabricated — the LLM is contracted to copy the returned string verbatim instead of constructing one itself. |\n| `check_claim_support` | Per-claim Sonnet judge: given a (claim, decision_id, optional pinpoint), return `supports: yes / partial / no / contradicts / unrelated` against the verbatim Erwägung text. Different model family (Sonnet) than the one running retrieval (Haiku) so retrieval errors are not re-introduced in verification. |\n| `attest_response` | Closing audit, called once before a final answer ships. Runs up to **five** rails over the LLM's draft:
① **case** — every BGE/BGer/BVGer/BStGer/BPatGer/MKGE reference exists, every pinpoint resolves;
② **statute** — every `Art. X LAW` reference resolves in `statutes.db`;
③ **quote** — every `\"…\"` of ≥30 chars appears verbatim in a cited source (decision or statute);
④ **date** — every `vom DD.MM.YYYY` adjacent to a citation matches the stored decision date;
⑤ **grounding** *(opt-in via `audit_grounding=true`)* — for each verified citation, an independent Sonnet judge checks whether the cited source actually supports the proposition the LLM attached to it.
Returns `linked_text` ready to ship verbatim with every validated citation wrapped in a Markdown link. |\n| `get_erwaegung` / `get_regeste` / `get_law` / `get_materialien` / `get_commentary` | The verbatim-text suppliers — the only sources the LLM is permitted to direct-quote. |\n\nThe architecture defends against **two empirically-measured legal-LLM failure classes**:\n\n- **Hallucination** (Dahl, Magesh, Suzgun & Ho, *Large Legal Fictions*, Stanford RegLab, 2024 — Journal of Legal Analysis): 58–88 % of legal queries to general-purpose LLMs produced at least one fabricated authority. A follow-up Stanford RegLab study of commercial legal-RAG tools (Magesh et al., *Hallucination-Free?*, 2024) measured 17–33 %. → caught by audits ① ② ③ ④.\n- **Reasoning error** (Butler & Butler, Isaacus, *Legal RAG Bench*, Mar 2026): citation is real and source was retrieved, but the proposition attached to it is not actually supported by the cited text. → caught by audit ⑤.\n\nThe full server prompt (R1–R8) embeds these rules so any connecting client (Claude, ChatGPT, Cursor, Gemini, Copilot Studio) inherits the contract automatically. See [`mcp_server.py`](https://github.com/jonashertner/caselaw-repo-1/blob/main/mcp_server.py) for the implementation, [`tests/web/test_attest_audits.py`](https://github.com/jonashertner/caselaw-repo-1/blob/main/tests/web/test_attest_audits.py) for the regression suite, and the [verification section on opencaselaw.ch](https://opencaselaw.ch/#verification) for a public explainer.\n\nEnd-to-end performance against this architecture is measured by **[Swiss Legal RAG Bench](https://github.com/jonashertner/caselaw-repo-1/blob/main/benchmarks/swiss_legal_rag_bench/README.md)** — a benchmark modelled on Butler & Butler's *Legal RAG Bench* methodology, covering DE/FR/IT federal-law questions and decomposing errors into hallucination / retrieval / reasoning components. v0.1 baseline (live OpenCaseLaw + Claude Sonnet 4.6, 10 questions): **100 % correctness, 90 % groundedness, 70 % retrieval accuracy**.\n\n### Example queries\n\nThese work on both the remote and local server:\n\n```\n> Search for BGer decisions on Mietrecht Kündigung from 2024\n\n> What did the BVGer rule on asylum seekers from Eritrea?\n\n> Show me the full text of 6B_1234/2023\n\n> How many decisions does each court in canton Zürich have?\n\n> Find decisions citing Art. 8 BV\n\n> What are the leading cases on Art. 8 EMRK?\n\n> Show me the citation network for BGE 138 III 374\n\n> How has Mietrecht jurisprudence evolved over time?\n\n> Show me Art. 41 OR (statute lookup)\n\n> Search for statute provisions about Verjährung\n\n> Trace the appeal chain for 5A_234/2026\n\n> Explain BGE 133 III 121 to me\n\n> What are the leading cases on Art. 41 OR and how has the doctrine evolved?\n\n> Give me a practice case on Haftpflichtrecht\n\n> Search for cantonal data protection laws\n\n> Show me the details of SR 220 (Obligationenrecht)\n\n> What legislation changed recently in Zürich?\n```\n\nThe AI calls the MCP tools automatically — you see the search results inline and can ask follow-up questions about specific decisions.\n\n### Citation graph tools\n\nFour tools expose the **reference graph**: 9.65 million resolved decision-to-decision citation edges (10 million extracted) and 12.4 million statute references. These require the graph database (`output/reference_graph.db`); if it's not available, the tools return a message instead of failing.\n\n**`find_citations`** — Given a decision, show its outgoing citations (what it references) and incoming citations (what references it). Each resolved citation includes the target decision's metadata and a confidence score. Unresolved references (e.g., older decisions not in the dataset) appear with their raw reference text.\n\n```\n> Show me the citation network for BGE 138 III 374\n\n## Outgoing (13 — what this decision cites)\n1. BGE 136 III 365 (2010-01-01) [bge] conf=0.98 mentions=1\n2. BGE 133 III 189 (2007-01-01) [bge] conf=0.93 mentions=1\n...\n\n## Incoming (13,621 — what cites this decision)\n1. 5A_234/2026 (2026-02-19) [bger] conf=0.92 mentions=2\n2. 5A_117/2026 (2026-01-30) [bger] conf=0.88 mentions=1\n...\n```\n\nParameters: `decision_id` (required), `direction` (both/outgoing/incoming), `min_confidence` (0–1, default 0.3), `limit` (default 50, max 200).\n\n**`find_appeal_chain`** — Trace the appeal chain (Instanzenzug) for a decision. Shows prior instances (lower courts) and subsequent instances (appeals to higher courts), reconstructing the full procedural path (e.g., Bezirksgericht → Obergericht → Bundesgericht).\n\nParameters: `decision_id` (required), `min_confidence` (0–1, default 0.3).\n\n**`find_leading_cases`** — Find the most-cited decisions, ranked by how many other decisions reference them. Filter by statute (law code + article), text query, court, or date range.\n\n```\n> What are the leading cases on Art. 8 EMRK?\n\n1. BGE 135 I 143 (2009) — 3,840 citations\n Regeste: Anspruch auf Aufenthaltsbewilligung...\n2. BGE 130 II 281 (2004) — 3,155 citations\n Regeste: Familiennachzug; gefestigtes Anwesenheitsrecht...\n```\n\nParameters: `query` (optional text), `law_code` + `article` (optional statute), `court`, `date_from`, `date_to`, `limit` (default 20, max 100). At least one of `query` or `law_code` is recommended; without any filter it returns the globally most-cited decisions.\n\n**`analyze_legal_trend`** — Year-by-year decision counts showing how jurisprudence on a topic or statute has evolved over time. Returns a table with counts and a visual bar chart.\n\n```\n> How has Art. 29 BV jurisprudence evolved?\n\n**Filter:** Art. 29 BV\n**Total:** 34,669 decisions\n\nYear Count Bar\n2000 275 █████\n2005 733 ████████████\n2010 1,028 █████████████████\n2015 1,536 ██████████████████████████\n2020 1,896 ████████████████████████████████\n2025 2,400 ████████████████████████████████████████\n```\n\nParameters: `query` (optional text), `law_code` + `article` (optional statute), `court`, `date_from`, `date_to`. At least one of `query` or `law_code` is required.\n\n### Statute lookup tools (local, federal + cantonal)\n\nTwo tools provide direct access to **Swiss law text** from the local mirror, covering both federal and cantonal jurisdictions with article-level FTS5 indexing and sub-millisecond lookup:\n\n- **Federal:** 5,525 laws / ~133,359 articles in each of DE/FR/IT from the [Fedlex](https://www.fedlex.admin.ch) SPARQL endpoint, mirrored monthly into `statutes.db`. Covers every consolidated federal act in German, French, and Italian.\n- **Cantonal:** 15,600 laws / 361,430 articles in `cantonal_laws.db`, sourced via two layers:\n - **Direct portal scraping (primary)** — 22 cantons whose official Gesetzessammlungen are published via the LexWork, SIL, ZH OpenData, or TI Raccolta delle Leggi platforms (AG, AI, AR, BE, BL, BS, FR, GE, GL, GR, LU, NE, NW, OW, SG, SH, SO, TG, TI, VS, ZG, ZH). The HTML is parsed natively — no PDF extraction, no OCR — yielding clean article-level data.\n - **LexFind PDF fallback** — 4 cantons not yet covered by direct scrapers (JU, SZ, UR, VD). PDFs from [LexFind.ch](https://www.lexfind.ch) are extracted with PyMuPDF and segmented into articles. Bilingual secondary-language passes for BE / FR / VS / GR.\n - Combined and federated via SQLite FTS5 with the federal table.\n\n**`get_law`** — Look up any Swiss law (federal or cantonal) by SR number or abbreviation, optionally fetching a specific article with full text.\n\n```\n> Show me Art. 8 BV\n\n# BV — SR 101\n**Bundesverfassung der Schweizerischen Eidgenossenschaft vom 18. April 1999**\n\n### Art. 8 — Rechtsgleichheit\n1 Alle Menschen sind vor dem Gesetz gleich.\n2 Niemand darf diskriminiert werden, namentlich nicht wegen der Herkunft, ...\n\n> Show me Art. 1 of the Zurich Hundegesetz\n\nget_law(canton=\"ZH\", sr_number=\"554.5\", article=\"1\")\n```\n\nParameters: `sr_number` or `abbreviation` (at least one required), `article` (optional — omit to see the full article list), `language` (de/fr/it, default de), `canton` (default `CH`; set to `ZH`, `BE`, etc. for cantonal lookup).\n\n**`search_laws`** — Unified FTS5 search across every Swiss statute article, federal and cantonal. BM25-ranked per corpus, merged by rank interleaving so each response surfaces both jurisdictions.\n\n```\n> Search for statute provisions about Verjährung\n\n1. [CH] Art. 130 OR (SR 220): Die Verjährung beginnt mit der Fälligkeit der Forderung...\n2. [ZH] § 19 Kirchgemeindenreglement (SR 182.60): ... Verjährungsbestimmungen ...\n3. [CH] Art. 132 OR (SR 220): Bei der Berechnung der Frist ist der Tag...\n```\n\nParameters: `query` (required, FTS5 syntax), `sr_number` (optional — one specific federal law, implies federal-only), `canton` (optional — restrict to one canton, or `CH` for federal-only), `jurisdiction` (`all` / `federal` / `cantonal`, default `all`), `language` (de/fr/it), `limit` (1–50).\n\n### Legislation discovery tools (LexFind-backed)\n\nThree additional tools layer the **live LexFind API** on top of the local mirror for cases where the mirror isn't enough — newly published versions between refresh cycles, full version history, recent-changes feeds. These are the broader discovery surface; for plain \"give me the current text\" queries `get_law` / `search_laws` are faster and more reliable.\n\n**`search_legislation`** — Broader discovery search across Swiss legislation, with optional single-call full-text enrichment. Useful when you don't know whether the law is federal or cantonal, or when you need a one-shot natural-language workflow.\n\n```\n> Search for the Bernese dog act with full text\n\nsearch_legislation(query=\"Hundegesetz\", canton=\"BE\", fetch_top_n_texts=2)\n→ returns top 2 matches with full_text_preview + sample_articles inline,\n no follow-up get_legislation call needed.\n```\n\nParameters: `query` (required), `canton` (optional — CH, ZH, BE, etc.), `active_only` (default true), `search_in_content` (default false — searches titles/keywords; set true to search law text), `language` (de/fr/it), `limit` (1–60, default 20), `fetch_top_n_texts` (0–10, default 0 — set to N to enrich top N results with parsed full article text).\n\n**`get_legislation`** — Get details for a specific law including metadata, version history, and links to official sources (Fedlex, cantonal portals).\n\n```\n> Show me the Obligationenrecht on LexFind\n\nSR 220 — Bundesgesetz betreffend die Ergänzung des Schweizerischen\nZivilgesetzbuches (Fünfter Teil: Obligationenrecht)\nEntity: Bund (CH) | Category: Gesetz | Keywords: OR\nIn force since: 01.01.2026\nSources:\n DE: https://www.fedlex.admin.ch/eli/cc/27/317_321_377/de\n FR: https://www.fedlex.admin.ch/eli/cc/27/317_321_377/fr\n```\n\nParameters: `lexfind_id` (from search results) or `systematic_number` + `canton` (e.g., \"220\" + \"CH\"), `include_versions` (default false), `language` (de/fr/it).\n\n**`browse_legislation_changes`** — Recent legislation changes for a canton or federal level. Shows new laws, amendments, and abrogations with dates and links.\n\n```\n> What legislation changed recently in Zürich?\n\n1. [01.03.2026] version — Vollzugsverordnung zur Finanzverordnung (SR 181.131)\n2. [01.03.2026] new — Studienordnung für den Masterstudiengang... (SR 414.253.125)\n3. [01.03.2026] version — Organisationsverordnung der Finanzdirektion (SR 172.110.3)\n```\n\nParameters: `canton` (default CH), `language` (de/fr/it).\n\n### Education tools\n\nThree tools support **legal study** covering the three core student workflows: understanding a case, understanding a doctrine, and practicing exam subsumption. Tools return rich structured data; Claude acts as the tutor and generates all pedagogy dynamically.\n\n**`get_case_brief`** — Any case reference (\"BGE 133 III 121\", docket number, or decision_id) → structured brief. Returns regeste, Sachverhalt, key numbered Erwägungen, Dispositiv, cited statutes with text excerpts, citation authority (incoming/outgoing count), and the top cited-by and cites cases.\n\n**`get_doctrine`** — Statute article (\"Art. 41 OR\") or legal concept (\"culpa in contrahendo\", \"Tierhalterhaftung\") → doctrine overview. Returns the leading cases ranked by citation count, a chronological doctrine timeline showing how rules evolved, and the statute text (if applicable).\n\n**`generate_exam_question`** — Legal topic (\"Haftpflichtrecht\", \"Art. 41 OR\", \"Mietrecht\") → real Fallbearbeitung. Picks a leading BGE, returns the anonymized Sachverhalt as fact pattern plus a hidden analysis (applicable statutes, legal test, correct outcome). The student writes a subsumption, then asks Claude to reveal the analysis for comparison and feedback. Pass `exclude_ids` to avoid repeating the same case.\n\n`draft_mock_decision` can use optional Fedlex URLs and caches fetched statute excerpts in\n`~/.swiss-caselaw/fedlex_cache.json` (configurable via `SWISS_CASELAW_FEDLEX_CACHE`).\n\n### Search quality benchmark\n\nUse a fixed golden query set to track search relevance over time:\n\n```bash\npython3 benchmarks/run_search_benchmark.py \\\n --db ~/.swiss-caselaw/decisions.db \\\n -k 10 \\\n --json-output benchmarks/latest_search_benchmark.json\n```\n\nMetrics: `MRR@k`, `Recall@k`, `nDCG@k`, `Hit@1`\n\nThe repository also ships a frozen offline baseline at `benchmarks/search_benchmark_2026-03-19_offline_full.json`. On the full 100-query set against a 1,078,177-row local `decisions.db`, that run recorded `MRR@10 = 0.4697`, `Recall@10 = 0.4958`, `nDCG@10 = 0.5250`, and `Hit@1 = 0.33`. Treat it as a reproducible offline baseline rather than a fully provisioned hosted-deployment score.\n\nYou can enforce minimum quality gates (non-zero exit on failure):\n\n```bash\npython3 benchmarks/run_search_benchmark.py \\\n --db ~/.swiss-caselaw/decisions.db \\\n -k 10 \\\n --min-mrr 0.50 \\\n --min-recall 0.75 \\\n --min-ndcg 0.85\n```\n\n### Build Reference Graph (Optional)\n\nFor statute/citation-aware reranking, build the local graph database:\n\n```bash\npython3 search_stack/build_reference_graph.py \\\n --source-db ~/.swiss-caselaw/decisions.db \\\n --courts bger,bge,bvger \\\n --db output/reference_graph.db\n```\n\nThen point the server to it:\n\n```bash\nexport SWISS_CASELAW_GRAPH_DB=output/reference_graph.db\n```\n\nGraph signals are enabled by default. To disable them, set `SWISS_CASELAW_GRAPH_SIGNALS=0`.\n\n### Build Statutes Database (Optional)\n\nFor statute lookup (`get_law`, `search_laws`), build the Fedlex statute database:\n\n```bash\n# Download the top 100 most-cited federal laws from Fedlex\npython3 -m scrapers.fedlex --top 100\n\n# Build the SQLite FTS5 statutes database\npython3 -m search_stack.build_statutes_db\n```\n\nThen copy to the data directory:\n\n```bash\ncp output/statutes.db ~/.swiss-caselaw/\n```\n\nOr set the path explicitly:\n\n```bash\nexport SWISS_CASELAW_STATUTES_DB=output/statutes.db\n```\n\n---\n\n## 2. Download the dataset\n\nThe full dataset is on [HuggingFace](https://huggingface.co/datasets/voilaj/swiss-caselaw) as Parquet files — one file per court, 34 fields per decision including complete decision text.\n\nMachine-consumable artifact metadata is published at `artifacts/manifest.json`. Besides daily Parquet/SQLite deltas, the manifest may point to an optional full compressed SQLite snapshot at `artifacts/sqlite/snapshots/.decisions.sqlite.zst` for tools that want to bootstrap a local FTS5 database without rebuilding from Parquet.\n\n### With Python (datasets library)\n\n**Step 1.** Install the library:\n\n```bash\npip install datasets\n```\n\n**Step 2.** Load the data:\n\n```python\nfrom datasets import load_dataset\n\n# Load a single court (~170k decisions, ~800 MB)\nbger = load_dataset(\"voilaj/swiss-caselaw\", data_files=\"data/bger.parquet\")\n\n# Load all courts (~900K decisions, ~6.5 GB download)\nds = load_dataset(\"voilaj/swiss-caselaw\", data_files=\"data/*.parquet\")\n```\n\n**Step 3.** Explore:\n\n```python\n# Print a single decision\ndecision = bger[\"train\"][0]\nprint(decision[\"docket_number\"]) # \"6B_1/2024\"\nprint(decision[\"decision_date\"]) # \"2024-03-15\"\nprint(decision[\"language\"]) # \"de\"\nprint(decision[\"regeste\"][:200]) # First 200 chars of the headnote\nprint(decision[\"full_text\"][:500]) # First 500 chars of the full text\n```\n\n### With pandas\n\n```python\nimport pandas as pd\n\n# Load one court\ndf = pd.read_parquet(\"hf://datasets/voilaj/swiss-caselaw/data/bger.parquet\")\n\n# Filter by date\ndf_recent = df[df[\"decision_date\"] >= \"2024-01-01\"]\nprint(f\"{len(df_recent)} decisions since 2024\")\n\n# Filter by language\ndf_french = df[df[\"language\"] == \"fr\"]\n\n# Group by legal area\ndf.groupby(\"legal_area\").size().sort_values(ascending=False).head(10)\n```\n\n### Direct download\n\nEvery court is a single Parquet file. Download directly:\n\n```\nhttps://huggingface.co/datasets/voilaj/swiss-caselaw/resolve/main/data/bger.parquet\nhttps://huggingface.co/datasets/voilaj/swiss-caselaw/resolve/main/data/bvger.parquet\nhttps://huggingface.co/datasets/voilaj/swiss-caselaw/resolve/main/data/zh_gerichte.parquet\n...\n```\n\nFull list of files: [huggingface.co/datasets/voilaj/swiss-caselaw/tree/main/data](https://huggingface.co/datasets/voilaj/swiss-caselaw/tree/main/data)\n\n---\n\n## 3. REST API\n\nQuery the dataset over HTTP without installing anything. This uses the [HuggingFace Datasets Server](https://huggingface.co/docs/datasets-server/).\n\n**Get rows:**\n\n```bash\ncurl \"https://datasets-server.huggingface.co/rows?dataset=voilaj/swiss-caselaw&config=default&split=train&offset=0&length=5\"\n```\n\n**Get dataset info:**\n\n```bash\ncurl \"https://datasets-server.huggingface.co/info?dataset=voilaj/swiss-caselaw\"\n```\n\n**Search by SQL** (DuckDB endpoint):\n\n```bash\ncurl -X POST \"https://datasets-server.huggingface.co/search?dataset=voilaj/swiss-caselaw&config=default&split=train\" \\\n -d '{\"query\": \"SELECT docket_number, decision_date, language FROM data WHERE court = '\\''bger'\\'' LIMIT 10\"}'\n```\n\n> Note: The REST API queries the dataset as configured in the HuggingFace repo (per-court Parquet files, full 34-field schema). For bulk access or local analysis, use the [download method](#2-download-the-dataset) above.\n\n---\n\n## 4. Web UI\n\nA local chat interface for searching Swiss court decisions. Ask questions in natural language, and an AI assistant searches the full corpus and answers with cited decisions.\n\n```\nBrowser (localhost:5173) → FastAPI backend → MCP server → Local SQLite FTS5 DB\n```\n\nEverything runs on your machine. No data leaves your computer (except LLM API calls to the provider you choose).\n\n### What you need\n\n| Requirement | How to check | Where to get it |\n|-------------|-------------|-----------------|\n| **Python 3.10+** | `python3 --version` (macOS/Linux) or `python --version` (Windows) | [python.org/downloads](https://www.python.org/downloads/) |\n| **Node.js 18+** | `node --version` | [nodejs.org](https://nodejs.org) — download the LTS version |\n| **An LLM provider** | *(see below)* | At least one cloud API key **or** a local model via Ollama |\n| **~65 GB free disk** | `df -h .` (macOS/Linux) | For the search index (downloaded on first run) |\n\n> **Windows users:** Install Python from [python.org](https://www.python.org/downloads/) and check \"Add Python to PATH\" during installation. Node.js installs npm automatically.\n\n#### Cloud providers (choose at least one, or use Ollama below)\n\n| Provider | Env variable | Where to get a key | Cost |\n|----------|-------------|---------------------|------|\n| **Google Gemini** | `GEMINI_API_KEY` | [aistudio.google.com/apikey](https://aistudio.google.com/apikey) | Free tier available |\n| **OpenAI** | `OPENAI_API_KEY` | [platform.openai.com/api-keys](https://platform.openai.com/api-keys) | Free credits for new accounts |\n| **Anthropic (Claude)** | `ANTHROPIC_API_KEY` | [console.anthropic.com](https://console.anthropic.com/) | Pay-as-you-go |\n\n> **Important:** A Claude Desktop or Claude Pro *subscription* does NOT include an API key. You need a separate developer account at [console.anthropic.com](https://console.anthropic.com/).\n\n#### Local models (no API key needed)\n\nIf you prefer not to use cloud APIs, you can run everything locally with [Ollama](https://ollama.com):\n\n| Model | Command to install | Download size | RAM needed |\n|-------|-------------------|---------------|------------|\n| Qwen 2.5 (14B) | `ollama pull qwen2.5:14b` | ~9 GB | ~16 GB |\n| Llama 3.3 (70B) | `ollama pull llama3.3:70b` | ~40 GB | ~48 GB |\n\nInstall Ollama from [ollama.com](https://ollama.com) (macOS, Linux, Windows), then:\n\n```bash\nollama serve # start the Ollama server (leave running)\nollama pull qwen2.5:14b # download a model (one-time)\n```\n\nThe Web UI auto-detects Ollama and shows local models as available.\n\n### Step-by-step setup\n\n**Step 1. Clone the repository:**\n\n```bash\ngit clone https://github.com/jonashertner/caselaw-repo-1.git\ncd caselaw-repo-1\n```\n\n**Step 2. Create a Python virtual environment:**\n\n```bash\npython3 -m venv .venv\n```\n\nActivate it:\n\n| OS | Command |\n|----|---------|\n| **macOS / Linux** | `source .venv/bin/activate` |\n| **Windows (PowerShell)** | `.venv\\Scripts\\Activate.ps1` |\n| **Windows (cmd)** | `.venv\\Scripts\\activate.bat` |\n\n> You'll know it's active when your terminal prompt starts with `(.venv)`.\n\n**Step 3. Install Python dependencies:**\n\n```bash\npip install fastapi uvicorn python-dotenv mcp pyarrow pydantic\n```\n\nThen install at least one LLM provider SDK:\n\n```bash\npip install anthropic # for Claude\npip install openai # for OpenAI / GPT-4o / local models via Ollama\npip install google-genai # for Google Gemini\n```\n\n> **Tip:** The `openai` package is also used for local Ollama models (Ollama exposes an OpenAI-compatible API). If you only want to use local models, `pip install openai` is sufficient — no cloud API key required.\n\n**Step 4. Install the frontend:**\n\n```bash\ncd web_ui && npm install && cd ..\n```\n\n**Step 5. Configure your API key:**\n\n```bash\ncp .env.example .env\n```\n\nOpen `.env` in a text editor and paste your API key on the appropriate line. For example, if you have a Gemini key, change `GEMINI_API_KEY=AI...` to your actual key. Leave the other provider lines as-is — they will be ignored if empty.\n\n> You can also skip this step and configure keys from the **Settings** panel inside the UI after starting.\n\n**Step 6. Start the app:**\n\n| OS | Command |\n|----|---------|\n| **macOS / Linux** | `./scripts/run_web_local.sh` |\n| **Windows (PowerShell)** | `.\\scripts\\run_web_local.ps1` |\n\nOpen **http://localhost:5173** in your browser.\n\n**What to expect on first run:** The MCP server will automatically download the dataset (~7 GB) from HuggingFace and build a local search index (~58 GB). This takes **30–60 minutes** depending on your connection and disk speed. You'll see progress in the terminal. After this one-time setup, the app starts instantly.\n\n### Features\n\n- **5 models**: Claude, OpenAI, Gemini (cloud) + Qwen 2.5 and Llama 3.3 via Ollama (local)\n- **Local-first option**: Run entirely on your machine with Ollama — no cloud API keys needed\n- **Streaming**: Responses appear token-by-token in real time\n- **Tool-augmented chat**: The AI calls search, get_decision, list_courts, etc. automatically\n- **Decision cards**: Clickable statute references with inline Fedlex article text\n- **Filters**: Narrow results by court, canton, language, and date range\n- **In-app settings**: Configure API keys and Ollama connection from the UI\n- **Export**: Download conversations as Markdown, Word, or PDF\n\n### Troubleshooting\n\n| Problem | Solution |\n|---------|----------|\n| `python3: command not found` (Windows) | Use `python` instead of `python3`, or reinstall Python with \"Add to PATH\" checked |\n| `npm: command not found` | Install Node.js from [nodejs.org](https://nodejs.org) |\n| `ModuleNotFoundError: No module named 'fastapi'` | Activate your venv (`source .venv/bin/activate`) and re-run `pip install ...` |\n| \"No provider configured\" banner | Click the gear icon (Settings) and paste an API key, or start Ollama |\n| \"Database not found\" on first run | Wait for the initial download to finish (check terminal for progress) |\n| Port already in use | Edit `BACKEND_PORT` or `FRONTEND_PORT` in `.env` |\n| PowerShell script blocked (Windows) | Run `Set-ExecutionPolicy -Scope CurrentUser RemoteSigned` once |\n\nFor advanced configuration (custom ports, MCP server path, timeouts), see [`.env.example`](https://github.com/jonashertner/caselaw-repo-1/blob/main/.env.example).\n\n---\n\n## What's in each decision\n\nEvery decision has 34 structured fields:\n\n### Core fields\n\n| Field | Type | Example | Description |\n|-------|------|---------|-------------|\n| `decision_id` | string | `bger_6B_1234_2025` | Unique key: `{court}_{docket_normalized}` |\n| `court` | string | `bger` | Court code ([full list](https://opencaselaw.ch)) |\n| `canton` | string | `CH` | `CH` for federal, `ZH`/`BE`/`GE`/... for cantonal |\n| `docket_number` | string | `6B_1234/2025` | Original case number as published |\n| `decision_date` | date | `2025-03-15` | Date the decision was rendered |\n| `language` | string | `de` | `de`, `fr`, `it`, or `rm` |\n| `full_text` | string | *(complete text)* | Full decision text, typically 5–50 pages |\n| `source_url` | string | `https://bger.ch/...` | Permanent link to the original |\n\n### Legal content\n\n| Field | Type | Description |\n|-------|------|-------------|\n| `regeste` | string | Legal headnote / summary (Regeste) |\n| `legal_area` | string | Area of law (Strafrecht, Zivilrecht, ...) |\n| `title` | string | Subject line (Gegenstand) |\n| `outcome` | string | Result: Gutheissung, Abweisung, Nichteintreten, ... |\n| `decision_type` | string | Type: Urteil, Beschluss, Verfügung, ... |\n| `cited_decisions` | string | JSON array of cited decision references |\n| `bge_reference` | string | BGE collection reference if published |\n| `abstract_de` | string | German abstract (primarily BGE) |\n| `abstract_fr` | string | French abstract |\n| `abstract_it` | string | Italian abstract |\n\n### Court metadata\n\n| Field | Type | Description |\n|-------|------|-------------|\n| `chamber` | string | Chamber (e.g., \"I. zivilrechtliche Abteilung\") |\n| `judges` | string | Panel composition |\n| `clerks` | string | Court clerks (Gerichtsschreiber) |\n| `collection` | string | Official collection reference |\n| `appeal_info` | string | Appeal status / subsequent proceedings |\n\n### Technical fields\n\n| Field | Type | Description |\n|-------|------|-------------|\n| `docket_number_2` | string | Secondary docket number |\n| `publication_date` | date | Date published online |\n| `pdf_url` | string | Direct URL to PDF |\n| `external_id` | string | Cross-reference ID |\n| `scraped_at` | datetime | When this decision was scraped |\n| `source` | string | Data source identifier |\n| `source_id` | string | Source-specific ID (e.g. Signatur) |\n| `source_spider` | string | Source spider/scraper name |\n| `content_hash` | string | MD5 of full_text for deduplication |\n| `has_full_text` | bool | Whether `full_text` is non-empty |\n| `text_length` | int | Character count of `full_text` |\n\nFull 34-field Parquet export schema: [`export_parquet.py`](https://github.com/jonashertner/caselaw-repo-1/blob/main/export_parquet.py)\n\n---\n\n## Coverage\n\n### Federal courts\n\n| Court | Code | Decisions | Period | Source |\n|-------|------|-----------|--------|--------|\n| Federal Supreme Court (BGer) | `bger` | ~173,000 | 1996–present | bger.ch |\n| BGE Leading Cases | `bge` | ~45,000 | 1954–present | bger.ch CLIR |\n| Federal Administrative Court (BVGer) | `bvger` | ~91,000 | 2007–present | bvger.ch |\n| Federal Admin. Practice (VPB) | `ch_vb` | ~23,000 | 1982–2016 | admin.ch |\n| Federal Criminal Court (BStGer) | `bstger` | ~11,000 | 2004–present | bstger.weblaw.ch |\n| EDÖB (Data Protection) | `edoeb` | ~1,200 | 1994–present | edoeb.admin.ch |\n| FINMA | `finma` | ~1,200 | 2008–2024 | finma.ch |\n| ECHR (Swiss cases, BGer-published) | `bge_egmr` | ~475 | 1974–present | bger.ch CLIR |\n| ECHR Switzerland (HUDOC) | `hudoc_ch` | 834 | 1959–present | HUDOC API |\n| ECtHR Chamber judgments | `ecthr_chamber` | 193 (growing) | 1959–present (FR; EN v2) | HUDOC API |\n| ECtHR Committee judgments | `ecthr_committee` | 30 (growing) | | HUDOC API |\n| ECtHR Grand Chamber | `ecthr_grand_chamber` | 13 (growing) | | HUDOC API |\n| Militärkassationsgericht (MKG) | `mkg` | 1,244 | 1915–2025 | oa.admin.ch + alexandria.ch |\n| Federal Patent Court (BPatGer) | `bpatger` | ~190 | 2012–present | bpatger.ch |\n| Competition Commission (WEKO) | `weko` | ~120 | 2009–present | weko.admin.ch |\n| Sports Tribunal | `ta_sst` | ~50 | 2024–present | ta-sst.ch |\n| Federal Council | `ch_bundesrat` | ~15 | 2012–present | bj.admin.ch |\n\n### Cantonal courts\n\n77 courts across all 26 cantons. The largest cantonal collections:\n\n| Canton | Courts | Decisions | Period |\n|--------|--------|-----------|--------|\n| Vaud (VD) | 3 | ~155,000 | 1984–present |\n| Zürich (ZH) | 20 | ~126,000 | 1980–present |\n| Genève (GE) | 1 | ~116,000 | 1993–present |\n| Ticino (TI) | 1 | ~58,000 | 1995–present |\n| St. Gallen (SG) | 7 | ~35,000 | 2001–present |\n| Graubünden (GR) | 1 | ~29,000 | 2002–present |\n| Basel-Landschaft (BL) | 1 | ~26,000 | 2000–present |\n| Bern (BE) | 6 | ~26,000 | 2002–present |\n| Aargau (AG) | 18 | ~21,000 | 1993–present |\n| Basel-Stadt (BS) | 3 | ~19,000 | 2001–present |\n\nAll 26 cantons covered: AG, AI, AR, BE, BL, BS, FR, GE, GL, GR, JU, LU, NE, NW, OW, SG, SH, SO, SZ, TG, TI, UR, VD, VS, ZG, ZH.\n\nLive per-court statistics: **[Dashboard](https://opencaselaw.ch)**\n\n---\n\n## How it works\n\n```\n ┌──────────────────────────────────────────────────────┐\n │ Daily Pipeline │\n │ │\nCourt websites ────────►│ Scrapers ──► JSONL ──┬──► Parquet ──► HuggingFace │\n bger.ch │ (45 scrapers, │ │\n bvger.ch │ rate-limited, ├──► FTS5 DB ───┐ │\n cantonal portals │ resumable) │ ├► MCP Server │\n │ └──► Graph DB ──┘ │\n │ ▲ │\nFedlex (SPARQL) ───────►│ Fedlex scraper ──► XML ──► Statutes DB ─┘ │\n │ │\n │ 01:00 UTC scrape 04:00 UTC publish │\n └──────────────────────────────────────────────────────┘\n```\n\n### Step by step\n\n1. **Scrape** (01:00 UTC daily) — 45 scrapers run in parallel, each targeting a specific court's website or API. Every scraper is rate-limited and resumable: it tracks which decisions it has already seen and only fetches new ones. Output: one JSONL file per court. A separate Fedlex scraper downloads federal law texts (Akoma Ntoso XML) via SPARQL for the statute database.\n\n2. **Build search index** (04:00 UTC) — JSONL files are ingested into a SQLite FTS5 database for full-text search. On Mon–Sat, this runs in **incremental mode**: a byte-offset checkpoint tracks how far each JSONL file has been read, so only newly appended decisions are processed (typically < 1 minute). On Sundays, a **full rebuild** compacts the FTS5 index and resets the checkpoint (~3 hours). Decisions appearing in multiple sources are deduplicated by `decision_id`, keeping the version with the longest full text. A quality enrichment step fills in missing titles, regestes, and content hashes.\n\n3. **Export** — JSONL files are converted to Parquet (one file per court) with a fixed 34-field schema.\n\n4. **Upload** — Parquet files are pushed to HuggingFace. The MCP server and `datasets` library pick up the new data automatically. Optional artifact publishing can also update `artifacts/manifest.json` with daily deltas and a full compressed SQLite snapshot for external bootstrap tools.\n\n5. **Update dashboard** — `stats.json` is regenerated (including scraper health status from the last run) and pushed to GitHub Pages.\n\n---\n\n## Running locally (developer)\n\nFor contributors and developers who want to run scrapers, build the pipeline, or modify the codebase.\n\n### Prerequisites\n\n- Python 3.10+\n- pip\n\n### Install\n\n```bash\ngit clone https://github.com/jonashertner/caselaw-repo-1.git\ncd caselaw-repo-1\npython3 -m venv .venv\nsource .venv/bin/activate # macOS/Linux — on Windows: .venv\\Scripts\\Activate.ps1\npip install -e \".[all]\"\n```\n\nThis installs all dependencies including PDF parsing, crypto, and the FastAPI server. For a minimal install without optional dependencies, use `pip install -e .` instead.\n\n### Scrape decisions\n\n```bash\n# Scrape 5 recent decisions from the Federal Supreme Court\npython run_scraper.py bger --max 5 -v\n\n# Scrape BVGer decisions since a specific date\npython run_scraper.py bvger --since 2025-01-01 --max 20 -v\n\n# Scrape a cantonal court\npython run_scraper.py zh_gerichte --max 10 -v\n```\n\nOutput is written to `output/decisions/{court}.jsonl` — one JSON object per line, one file per court. The scraper remembers what it has already fetched (state stored in `state/`), so you can run it repeatedly to get only new decisions.\n\n45 court codes are available. Run `python run_scraper.py --list` for the full list, or see the [dashboard](https://opencaselaw.ch) for per-court statistics.\n\n### Build a local search database\n\n```bash\n# Full build (reads all JSONL, optimizes FTS index — ~3h for 900K decisions)\npython build_fts5.py --output output -v\n\n# Incremental build (reads only new JSONL bytes, skips optimize — seconds)\npython build_fts5.py --output output --incremental --no-optimize -v\n\n# Full rebuild (deletes DB + checkpoint, rebuilds from scratch)\npython build_fts5.py --output output --full-rebuild -v\n```\n\nThis reads JSONL files from `output/decisions/` and builds a SQLite FTS5 database at `output/decisions.db`. A full build of 900K decisions takes about 3 hours and produces a ~58 GB database. Incremental mode uses a checkpoint file (`output/.fts5_checkpoint.json`) to skip unchanged files and seek past already-processed bytes, completing in seconds when few new decisions exist.\n\n### Export to Parquet\n\n```bash\npython export_parquet.py --input output/decisions --output output/dataset -v\n```\n\nConverts JSONL files to Parquet format (one file per court). Output goes to `output/dataset/`.\n\n---\n\n## Data sources\n\n| Source | What | How |\n|--------|------|-----|\n| **Official court websites** | Federal courts (bger.ch, bvger.ch, bstger.ch, bpatger.ch) | JSON APIs, structured HTML |\n| **Federal regulatory bodies** | FINMA, WEKO, EDÖB, VPB | Sitecore/custom APIs |\n| **Cantonal court portals** | 26 cantonal platforms (Weblaw, Tribuna, FindInfo, custom portals) | Court-specific scrapers |\n\nDecisions appearing in multiple sources are deduplicated by `decision_id` (a deterministic hash of court code + normalized docket number). The version with the longest full text is kept.\n\n---\n\n## Legal basis\n\nCourt decisions are public records under Swiss law. Article 27 BGG requires the Federal Supreme Court to publish its decisions. The Bundesgericht has consistently held that court decisions must be made accessible to the public (BGE 133 I 106, BGE 139 I 129). This project scrapes only publicly available, officially published decisions.\n\n## Governance and removal\n\nRepublication changes discoverability, so the project ships a governance policy covering source withdrawals, re-anonymization, and verified correction/removal requests. See [docs/governance-and-removal-policy.md](https://github.com/jonashertner/caselaw-repo-1/blob/main/Users/jonashertner/caselaw-repo-1/docs/governance-and-removal-policy.md).\n\n---\n\n## Contributors\n\nMaintainer: [Jonas Hertner](https://jonashertner.com). A small group of first users, bug reporters, institutional partners, and consumer-side integrators have shaped the tool; see [CONTRIBUTORS.md](https://github.com/jonashertner/caselaw-repo-1/blob/main/CONTRIBUTORS.md) for the full list and how to add yours.\n\nIf you're using OpenCaseLaw in production — as a scraper contributor, dataset consumer, law-firm integrator, or research project — we'd like to know. Email **team@jonashertner.com** or open a [discussion](https://github.com/jonashertner/caselaw-repo-1/discussions).\n\n---\n\n## License\n\nCode: MIT. See [LICENSE](https://github.com/jonashertner/caselaw-repo-1/blob/main/LICENSE).\n\nDataset packaging and added metadata: CC0-1.0 to the extent rights exist. The underlying decision texts remain official published court decisions sourced from the originating courts and publ","changelogMarkdown","videoKey","requiredSecretsJson","publishedAt","priceCents","trialDays",0,"iconType","image","upstream",{"_109":110,"_111":112,"_113":114,"_115":116,"_117":118,"_119":120},"requiredSecrets",[],"tags",[100,101,102,103,104,105,106,107,108],"assets",[],"stats",{"_98":85,"_99":85},"ratingAvg","reviewCount","ai-agents","caselaw","citation-graph","dataset","legal-data","legal-research","open-data","swiss-law","switzerland","topics",[100,101,102,103,104,105,121,122,123,106,107,108],"license","MIT","language","Python","registryName","ch.opencaselaw/swiss-caselaw","homepage","https://opencaselaw.ch","packages",[],"mcp","mcp-server","model-context-protocol","ENV",{"_126":127,"_128":129,"_130":131},"ENVIRONMENT","production","TURNSTILE_SITE_KEY","0x4AAAAAAB8-rEI2gNDf_QhU","SENTRY_DSN","https://45fd1229f78089d5c3ee78be128030e1@o4510583261757440.ingest.de.sentry.io/4510584253448272"]

OpenCaseLaw — Swiss Caselaw, Statutes & Doctrine