🔭 Deep Research Canvas

Ask a research question. Watch it become a board of cited, connected data cards — converging on a verdict.

---

Why it's powerful

Most AI answers are a wall of text you just have to trust. Deep Research Canvas turns a single question into a spatial board you can see, navigate, and interrogate:

• 🧩 A living canvas, not a chat log. Every finding is a draggable card; edges show how sub-questions feed the consensus. Pan, zoom, select a node and ask about that node in the side chat.
• 📌 Grounded in real, cited data. With the tako provider, cards are backed by live Tako knowledge cards — real numbers, real sources, an "as-of" date — not the model's memory.
• ⚖️ See the difference, side by side. Flip the provider to gpt or claude (baselines, no tools) and the board goes thin and flagged model · <confidence> with no sources. Grounded boards are dense with citations. The gap is the whole point.
• 🎯 A deterministic verdict. The consensus / leaderboard ranking is plain app code over your stated criteria — reproducible, not vibes.

Quick start

git clone https://github.com/25eliu/deep-research-canvas.git
cd deep-research-canvas
npm install
cp .env.example .env.local     # add your API keys (see below)
npm run dev                    # → http://localhost:3000

Then ask something like "Research the best 5 semiconductor companies to invest in", switch providers in the top-left, and select nodes to dig deeper in the side chat.

Keys you'll need

Everything runs live and server-side — keys never touch the browser. Fill these into .env.local:

Variable	Powers	Where to get it
TAKO_API_KEY	Grounded, cited data cards (the tako provider)	developer.tako.com · docs.tako.com
OPENAI_API_KEY	The gpt baseline	platform.openai.com
ANTHROPIC_API_KEY	The claude baseline	console.anthropic.com

TAKO_HOST defaults to https://tako.com (production) — no need to set it. Set OPENAI_MODEL / ANTHROPIC_MODEL to current model strings for your accounts.

Run the tests with npm test.

How it works

The whole app is one idea: the agent is a pure function — (message + canvas state + selection + provider) → { canvasOps, narration, sideReply } — and the frontend just applies the ops. Switching providers changes only the grounding step; the decision logic stays identical.

• gpt / claude answer from memory and draw their own charts — no tools, no sources.
• tako runs a graph-first pipeline (resolve entities/metrics → fetch real cards → deterministic consensus), with grounded follow-ups via Tako Answer.

That's the overview — the full decision trees and Mermaid diagrams live in docs/agents-architecture.md. Built with Next.js, the Vercel AI SDK + Zod, and Tako for grounding.

License

MIT — see LICENSE. Not investment advice; the consensus is an analytical ranking on stated criteria.