DataRaum

The understanding layer that grounds an organization's operating model in its own data.

A semantic layer tells BI tools what columns are called. DataRaum learns what they mean — the concepts, relationships, rules, and measures of the organization — and grounds each one in the actual data, with a measured confidence behind it. See the docs for the full picture.

Monorepo layout

packages/
├── engine/          # Python — pipeline, detectors, Temporal activity worker
├── cockpit/         # TypeScript — TanStack Start web UI
├── dataraum-config/ # YAML data — entropy config, LLM prompts, verticals (bind-mounted, never imported)
└── infra/           # docker-compose orchestration

Each package has its own README. Start there if you're working in a specific package.

Status

DataRaum runs as a multi-container platform, isolated per workspace:

• engine (Python) — a Temporal activity worker, no HTTP. Does the durable analysis (add_source, begin_session, operating_model) and writes metadata to the workspace's Postgres schema.
• cockpit (TanStack Start) — the web app you use. Hosts the chat agent, renders the results, and orchestrates the journey by triggering engine workflows via Temporal.

They share one substrate: Postgres (metadata + cockpit state + catalogs), an S3 object store (the DuckLake data lake + uploads), and Temporal (durable orchestration). No HTTP seam between engine and cockpit — the integration surface is Postgres + Temporal. See the platform architecture.

Quick start

# Set the LLM key
cp packages/infra/.env.example packages/infra/.env
echo "ANTHROPIC_API_KEY=sk-ant-..." >> packages/infra/.env

# Bring up the full stack (Postgres, object store, Temporal, engine worker, cockpit)
docker compose -f packages/infra/docker-compose.yml up -d --wait

# Engine health = the Temporal worker heartbeat (no HTTP endpoint):
docker compose -f packages/infra/docker-compose.yml run --rm --no-deps \
  --entrypoint temporal temporal-admin-tools \
  worker list --namespace default --address temporal:7233   # → Status: Running

# Open the cockpit
open http://localhost:3000

For UI iteration, run the cockpit dev server outside docker for hot reload — see packages/cockpit/README.md.

Develop

• Engine (Python): cd packages/engine && uv sync --group dev && uv run pytest --testmon tests/unit -q. See packages/engine/README.md and packages/engine/CLAUDE.md.
• Cockpit (TypeScript): cd packages/cockpit && bun install && bun run dev. See packages/cockpit/README.md and packages/cockpit/CLAUDE.md.
• Pull the engine metadata schema (cockpit): cd packages/cockpit && DATARAUM_WORKSPACE_ID=<id> METADATA_DATABASE_URL=<url> bun run db:pull:metadata. Re-run after the engine adds/changes SQLAlchemy models.

Documentation

Platform docs live in docs/ (workspace root) and are published via Zensical. Start at docs/index.md, or serve the site locally:

uv run --project packages/engine zensical serve   # run from the repo root

• Overview — what DataRaum does, at a glance
• The approach · the journey · pipeline & phases · learnable surface · measurement & detectors
• Platform architecture — under the hood
• Decision records — the why behind the architecture

License

MIT — see LICENSE.