Document Workbench — Context Hitch

Most AI products that offer to “read your documents” do something more modest than that phrase suggests. You upload a PDF. You ask a question. A confident answer appears. The model did not read your PDF — it received a handful of passages chosen by machinery you never saw, and those passages are what it answered from. Whether the answer is any good depends on the machinery as much as it depends on the model.

The Document Workbench is the same kind of tool with the machinery exposed. Pick a document. Ask a question. You will see the answer, the specific passages the model was given, how many tokens were sent, and what the round-trip cost. When the answer is wrong, you will be able to see why it was wrong.

This is a teaching tool. It is not a replacement for Lexis, Westlaw, PubMed, or any other professional research platform. It is not a substitute for your own judgment about the documents that actually matter to your work.

Try one we’ve prepared

A small library of documents chosen because they teach something. Regulatory guidance on AI use — NIST, plus state-issued guidance from the Nevada Division of Insurance and the New York Department of Financial Services. Landmark cases every lawyer recognizes — Miranda, Chevron, and Hawkins v. McGee (the latter from the New Hampshire Supreme Court, which itself teaches something about the workbench’s reach across federal and state primary law). Famous AI failures, including the court order that sanctioned a lawyer for filing fake citations. A few public-domain long works for when you want to see what AI can and cannot do with eight hundred thousand tokens of Tolstoy.

Pick one. Dot will load it into the workbench alongside you and get out of the way.

Loading catalog…

Upload your own

Drop in a PDF, Word document, or plain text file up to 25 MB. Text only — no image-only PDFs, no scans. The workbench pre-translates it while you wait (typically a few seconds), and then you can ask it questions the same way you would one of ours.

Do not upload confidential material. Not because we are looking at it — we are not — but because this is a demonstration tool running on shared infrastructure. If you need to do serious work with AI on a document you have a duty to protect, use a properly configured API endpoint or a tool your employer has evaluated. This site is not that tool.

Documents you upload disappear when your session ends, or after 24 hours of inactivity, whichever comes first. There is no “recover my upload” button, because there is nothing stored to recover.

Upload track coming soon.

Try a guided lesson

Scattered through the prepared-document library are short lessons built around specific documents — Why the same question gets different answers, Check the AI’s work, When was this written?, and others. Each lesson walks you through a specific thing most professionals misunderstand about how AI reads documents, using a document you can verify against. You do not need to finish one to start another. You do not need to finish any to use the rest of the workbench.

Guided lessons coming soon.

Dot

Dot is the AI guide persona on this module. She answers your questions about the document you’ve loaded, shows her work, and occasionally steps back to observe something about the exchange — a leading question, a stretch in her own answer, a place where your prior knowledge is more reliable than hers. These observer callouts appear separately from her primary answers so you can see the difference.

Dot is an instructed instance of Claude. She is not a different AI. She is Claude, shaped by a specific system prompt and given access to specific retrieval machinery. When she makes a technical claim about how AI systems work, she is instructed to cite a source or say that she can’t. When retrieval fails, she is instructed to decline rather than make something up. Both behaviors are part of what this module is teaching.

If you want to know exactly what models she uses, what retrieval parameters, what is and is not logged, and what it costs to run her — that information is on the How This Works page, which is also the document Dot reads when you ask her what she is.

Chunks sent to Dot are roughly 564 tokens mean (NIST AI RMF 1.0 baseline). By default five chunks are sent to Dot; ranks 6–10 are visible in the right panel as retrieved-but-cut, so you can see what she didn’t see.

What The Workbench Is Not

A legal, medical, financial, or professional research tool. The documents in the prepared library are teaching artifacts. Do not cite them in a brief. Verify anything important against primary sources and the tools you are licensed to rely on.
A persistent workspace. Every visit is a new session. Nothing carries over. Your last session’s documents are gone.
A confidential environment. See above.
A general-purpose chatbot. Dot answers about the document you’ve loaded. She will not write your wedding toast.

A Note On Cost

Every question on the workbench costs a small amount of real money — typically a cent or two, occasionally more in full-context mode. The module displays the cost of each call as it happens. This is not a gimmick. It is part of the curriculum. Professional AI tools have prices; understanding those prices is part of using them responsibly.

The site has a daily spending cap. When it is hit, the AI features pause until the next day. The static reading material stays available. This is the honest answer to “what happens when this gets popular.”