Ursa

Database / data-access layer for Constellation’s research stack.

Ursa is a multimodal time-series database. It owns ingestion, indexing, lazy time-windowed reads, vector search, and lifecycle (retention, GC, Polaris cache sync) for all of Constellation’s multimodal data — EEG, video, eye tracking, biometrics, questionnaires, keyboard/mouse/screen captures, and more.

The data model is intentionally minimal: Participant → Recording (recording_hash) → Modality + Events + flexible metadata. No sessions, no trials, no stimuli.

Cloudflare R2 storage is backed by a Lance metadata catalog and Zarr time-series store; the user-facing API is Pydantic-typed and built on temporaldata primitives.

Contents

• Architecture
  • Goals
  • Data model
  • Catalog tables (Lance)
  • Storage backends
  • Lifecycle
  • Phasing
• Concepts
• Quickstart
• Tutorials
• API Reference

Where this fits

Ursa is one of three packages in Constellation’s research stack:

• Ursa (this site) — database / data layer

• Virgo — DAG-based preprocessing

• Orion — research / training / benchmarking

Full architecture: Research Stack Architecture (Notion).

Status

🌱 Early bootstrap. Implementation tracked in the Linear Ursa project.

Phasing (mirrors the Linear project milestones):

• M1 — Foundations (in progress)

• M2 — Underbuilt MVP (Phase 1a, solo)

• M3 — Production Database (Phase 1b, with Killian)

• M4 — Production-scale (Phase 4)

• M5 — Benchmarks Integration (Phase 5)

• M6 — Polish & Onboarding (Phase 6)