NewMulti-doc workspace · Notepad · Compare

From papers
to understanding
your field.

Every paper becomes a knowledge map. Cite findings, compare sources, export your review.

1–3 PDFs———→Knowledge Graphs———→Insights———→Chat———→Notepad———→Report
research-canvas/attention-is-all-you-need.pdf
12 nodes14 edges5 insightslive
Document1,847 words

Recurrent neural networks, long short-term memory and gated recurrent neural networks in particular, have been firmly established as state of the art approaches in sequence modeling…

We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.

Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU.

On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.0 after training for 3.5 days on eight GPUs…

◆ Focus"Transformer, based solely on attention…"
Knowledge Graphdagre · TB
based onusesachievesimplementsENTITYTransformerMETHODSelf-AttentionCONCEPTSequence ModelingFINDING28.4 BLEU scoreMETHODMulti-Head Attention
concept
entity
method
finding
Findinghigh

Self-attention beats recurrence

Methodologyhigh

Multi-head attention layers

Implicationmedium

Less training time at scale

Gaplow

Audio & vision modalities

Why does attention scale better?Send

Click → Highlight

Click any node — the verbatim source sentence glows in the document pane on the left.

Parallel agents

Graph extraction and insight synthesis run as concurrent Gemini calls — both bars fill at once.

Grounded chat

Every chat turn ships full DOCUMENT + GRAPH + FOCUS context to the model.

Capabilities

A research hub, not a chatbot.

Six primitives built for literature review.

Multi-doc workspace

Three papers, one canvas.

Load up to 3 PDFs simultaneously. Each gets its own colour-coded tab, persistent knowledge graph, insights panel, and chat history — all in a single workspace.

Bidirectional highlight

Click a node, the sentence glows.

Every graph node carries the verbatim source quote. Clicking finds and highlights the exact passage in the document — zero round-trips.

📓Research notepad

Cite anything, export everything.

Hit Cite on any graph node, insight card, or chat reply. Citations land in a persistent notepad with inline annotations and tags. One click exports a structured research report.

In-canvas compare

Side-by-side, no page switch.

Toggle Compare mode to see two documents, their graphs, and chats side-by-side. Session state is never lost — it's a layout toggle, not a new route.

Pipeline

Four steps, ~20 seconds.

From raw papers to a citable research report.

01Drop

Upload 1–3 PDFs

Drag papers onto the canvas tabs. Each gets its own colour-coded slot. PDF text is extracted in the browser — no server upload.

02Analyze

Two agents per doc

Graph Extractor builds 8–20 typed nodes. Insight Analyzer ranks structured insights. Both run in parallel from Gemini 2.5 Flash.

03Explore

Click · Cite · Compare

Click nodes to highlight source sentences. Hit Cite on anything to add it to the Notepad. Switch to Compare mode to diff two papers side-by-side.

04Export

Report from your citations

Annotate and tag each citation in the Notepad. Export a structured Markdown or PDF report built from your cited findings — not a raw dump.

Architecture

Four specialized agents.

Every prompt is hand-tuned, JSON-mode locked, Zod-validated, and runs per document slot.

A
Agent A

Graph Extractor

Builds a typed knowledge graph per slot

Returns 8–20 nodes across nine categories (concept, entity, method, finding, dataset, metric, result, assumption, limitation) with verbatim source quotes. Runs independently per document slot.

Output{ nodes: [...], edges: [...] }
B
Agent B

Insight Analyzer

Distills each paper into ranked insights

Up to 12 insights ranked by significance, tagged by category (finding · limitation · methodology · implication · gap · contribution) and confidence. Runs in parallel with Agent A.

Output{ insights: [...], summary }
C
Agent C

Context Chat

Streams answers grounded in the active doc

Receives DOCUMENT + GRAPH + FOCUS quote on every turn. References graph nodes inline as clickable chips. Each document slot has its own independent chat history.

Outputstream → ReadableStream
D
Agent D

Notepad + Report

Turns citations into a structured report

Cite any node, insight, or chat reply. Annotate and tag citations in the persistent Notepad sidebar. Export builds a grouped Markdown or PDF report from your cited findings — not a raw dump.

Output.md / .pdf report
Design system · Obsidian Lab

Not a chatbot.
A research instrument.

Inspired by Nature journal and terminal interfaces. Every element earns its place. Three-pane workspace, persistent per-document state, a Notepad that bridges exploration to writing, and Compare mode that never loses your session. Built for researchers doing real literature review.

Concept#E8A231
Entity#4A9EFF
Method#3ECF8E
Finding#E85D82

Aa

Instrument Serif

Aa

JetBrains Mono

Aa

Geist Sans

◆ Ready to read differently

Open the canvas.
Drop your papers.

No signup. No database. Load up to 3 papers, explore their graphs, cite findings to your Notepad, and export a structured report.