RabbitHole

Go down the research rabbit hole — and keep the map.

RabbitHole turns a 2am Wikipedia spiral into a structure you can navigate. You start with a question, then pull threads to dig deeper. Each pull fetches real findings from the web and nests them under the thing you pulled from, so the board grows into a branching tree you built by choosing which threads to yank — and you can always trace your way back.

Live at rabbithole.app.space.

Two views of one tree

There is a single underlying structure — the investigation tree of nested findings — rendered two ways, toggled by a view switch:

• Outline (default) — a nestable, collapsible threaded outline. The root is your question; every finding is a node with a source lens and a "pull this thread" action. Indentation and connector lines show the branch structure.
• Columns — a Miller-column / Finder-style drill-down. Selecting a node opens a new column to its right showing its children; you march rightward as you go deeper, and the columns behind you form the trail.

Both read and write the same tree — they're different renderings, not separate features.

The lenses

Every finding is tagged with the source lens it came from, color-coded and shown as a small chip. You choose which lens to pull a thread with:

• Exa — the deep neural / citation trail (default, primary engine)
• Wikipedia — the established, reference account
• News — what's being reported right now
• Web — the broad sweep

Following the trail

• A breadcrumb strip showing the path from the root to where you are.
• A trail panel of where you've been.
• Backlinks — when a new finding shares a source with an existing node, the connection is surfaced ("N findings mention this") instead of silently duplicating.

Every node carries its real, clickable sources. The trail is always backed by links.

The telling dial

A slider controls how each finding is narrated, from Straight (plain and factual) to Deep end (more speculative, connect-the-dots framing). It only changes narration style — the underlying sources are always real and cited.

How it's built

A DeepSpace app on Cloudflare Workers.

• Data — two collections (holes, nodes). The tree is a flat node list linked by parentId; findings are written server-side (server actions) so they're durable and immediately consistent for every client.
• Digging — a pullThread server action queries the chosen lens (exa / wikipedia / newsapi / websearch) with context from the parent node, synthesizes each result into a short narrated finding (respecting the tone dial), and keeps the real source URL attached. Backlinks are computed by shared source URL.
• Sharing — a hole is private until its owner shares it; then anyone with the link can follow the trail, read-only. The model leaves room for forks, a public gallery, and real-time multiplayer without reshaping the tree — none of which are built yet.

Develop

npx deepspace dev       # local dev (vite + worker)
npx deepspace test      # smoke + api specs
npx deepspace deploy    # deploy to rabbithole.app.space