ReCan Flow
ReCan Flow is a local-first visual AI research workspace. It turns a question into a spatial research graph: decompose it into branches, research each branch with AI, save findings, synthesize findings into an insight, critically challenge that insight, and spin up a new question from the challenge — repeating the loop.
The research loop
Question → AI decomposition → Branches → Research (AI chat) → Findings
→ (select 2+) Synthesis → Insight → Challenge → New Question → …
- Question — the root of a research thread. "Explore with AI" decomposes it into 3–5 research branches.
- Branch — a research direction proposed by AI. "Explore" opens (or creates) its Research node.
- Research — a contextual AI chat scoped to one branch. Any assistant reply can be saved as a Finding.
- Finding — a saved piece of evidence with provenance back to its question/branch/research.
- Insight — created by selecting 2+ Findings and synthesizing them; carries a confidence level.
- Challenge — critically stress-tests an Insight (weaknesses, missing evidence, alternative explanation) and can spawn a new Question from its suggested follow-up.
- Note / Text — freeform annotations, no AI involved.
Architecture
- Canvas:
@xyflow/react, one node component per kind under components/workspace/nodes/. Layout for AI-generated children (branches, findings, insights, new questions) is deterministic and overlap-avoiding (lib/layout.ts), not random.
- State: a single
nodes/edges pair in components/workspace/Workspace.tsx. Node data holds AI status/error inline (no separate loading state tree). Selection lives on XYFlow's own node.selected/edge.selected fields rather than duplicated state, so click-driven and programmatic selection never disagree.
- Undo/redo:
lib/history.ts — a debounced snapshot stack so a drag or a burst of edits becomes one undo step.
- AI:
app/api/ai/route.ts is the only place that talks to a model. Four actions — decompose, chat, synthesize, challenge — each validated with Zod on the way in and the way the model's JSON comes back out. Gemini is primary, Groq is the fallback; if neither key is configured the route returns a clearly-labeled local placeholder rather than pretending research happened. Nothing here ever claims to have browsed the web or fabricates a citation.
- Persistence: versioned
localStorage (lib/workspace.ts). Corrupt or old-schema data is discarded in favor of a fresh empty workspace rather than crashing.
Local setup
pnpm install
copy .env.example .env.local
pnpm dev
Set GEMINI_API_KEY and/or GROQ_API_KEY in .env.local. Without either, AI actions return a local, clearly-labeled fallback instead of a real response. Never expose either key through a NEXT_PUBLIC_ variable.
Verification
pnpm typecheck
pnpm lint
pnpm build
Keyboard shortcuts
V select · H hand/pan (also Space+drag) · Q new question · N new note · R new research (or open the selected branch's) · F save the open research reply as a finding · I synthesize 2+ selected findings · Delete/Backspace remove selection · Ctrl/Cmd+Z undo · Ctrl/Cmd+Shift+Z redo · Ctrl/Cmd+K search nodes · Escape clear selection / close overlays.
Current limitations
- No grounded web search: the AI reasons from model knowledge only, and the UI never claims otherwise.
- Single local workspace per browser (no accounts, no multi-workspace switching, no real-time collaboration — Share just copies the URL).
- Desktop-first (1280px+); usable but not polished below ~800px.
