DOCUMENTS → GRAPH → UNDERSTANDING

Turn documents
into a living map.

An evolving knowledge workspace that extracts concepts, preserves evidence, and lets you project the same understanding into Obsidian.

Open the workbench ↓ See a real graph → See the projection →
0 concepts 0 relations 0 source docs
FIRST PRINCIPLES✳RUNNABLE EXPERIMENTS✳HONEST TRADE-OFFS✳SHARE THE WORK✳ FIRST PRINCIPLES✳RUNNABLE EXPERIMENTS✳HONEST TRADE-OFFS✳SHARE THE WORK✳

Give it a document.
Keep the structure.

By default, extraction runs in this browser. If you configure a same-origin model endpoint, the graph is still yours, the evidence stays attached, and every pass becomes a revision you can inspect or export.

02 / INTERNAL REPRESENTATIONREV 000
No concepts yet — ingest a document to begin.
+ Add manual knowledge
⌁

Your graph will appear here.

Nodes are concepts. Lines are evidence.
concept source inferred relation
03 / EXTERNAL PROJECTIONOBSIDIAN-READY

Same graph.
Different surface.

Export a navigable Markdown projection with YAML frontmatter, backlinks, relation lists, and source evidence. Open it in Obsidian, open the included native Graph.canvas view, edit the notes, and bring corrections back as feedback.

Full exports include source text, evidence quotes, and source URIs. Use the redacted actions for public sharing.

_knowledge-graph.mdpreview
Ingest a document to generate an Obsidian projection.

Less mystique.
More contact with reality.

A document is useful once its ideas can be found, connected, and challenged. The workbench turns unstructured text into a transparent internal representation instead of hiding the path behind a single answer.

The loop is: ingest → infer → inspect → improve → project. The graph is not the final truth. It is a living hypothesis with evidence attached and a revision history.

Pick a thread.
Follow it all the way down.

Start anywhere. The map is arranged by dependency, not prestige. The best next note is usually the one that makes your current project less mysterious.

⌕ /

One month.
One working artifact.

A deliberately small syllabus for building intuition through contact. You do not need to finish every exercise. You do need to keep making things.

FIELD
LAB ✳

Train a tiny
character model.

Take a text file you care about. Build a tokenizer. Train a small transformer. Watch the loss curve move. Then make it fail in three interesting ways.

DIFFICULTY ★★☆ TIME 2–4 evenings OUTPUT a notebook + model
See the build brief ↗
train.py
# the loop is the lesson
for step in range(max_steps):
    xb, yb = get_batch()
    logits, loss = model(xb, yb)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

print(f"step {step}: {loss.item():.3f}")
loss is falling ▁▂▃▅▇

Make the map
worth trusting.

These are not commandments. They are useful constraints for a fast-moving field.

01

Show your work.

Primary sources, runnable code, screenshots of failure. Make claims easy to inspect.

02

Keep it small.

The smallest useful explanation beats the grand unified theory every time.

03

Name the edges.

Performance, cost, latency, data, and weird behavior are part of the concept.

04

Leave a handle.

Every page should give the next curious person somewhere concrete to start.

Found a gap?
Pull the thread.

The best version of this resource cannot be written by one person. Add a correction, a sharper analogy, a benchmark, a failed attempt, or a link that deserves more context.

THE SHORT VERSION

BUILD IT

NEXT EXPERIMENT