How to upload documents and files to Whizi

Supported file types, and what actually happens to each

Attaching a file is one action, but what the model receives differs by type, and knowing which is which explains most of the surprising results.

Type	Extensions	What the model works with
Text PDFs	.pdf	The extracted text, in reading order, including tables as text
Scanned PDFs	.pdf	Text recovered by OCR, which is very good but not perfect
Word documents	.docx	The document text; comments and tracked changes may not carry through
Spreadsheets	.xlsx, .csv	Cell values as text, not live formulas
Images	.jpg, .png, .webp	The image itself, read by a multimodal model
Code and plain text	.py, .js, .ts, .go, .md, .txt and similar	The file contents verbatim

Two consequences worth internalising. First, a spreadsheet arrives as values rather than as a working workbook, so the model can read what your formula produced but cannot re-execute the formula. Second, a scanned PDF has passed through OCR, which means an unusual font, a handwritten annotation, or a low resolution scan can produce a misread character in a number. Verify figures from scanned documents against the original.

Prepare the file before you ask anything

Five minutes here saves a confusing conversation later.

• Name the file for what it is. The model sees the filename, and 2025-q3-vendor-contract-acme.pdf gives it useful context that document(3).pdf does not. It also matters enormously once several files are in the thread.
• Send the source, not the presentation version. For spreadsheets especially, the version with merged cells, subtotals inside the data, and notes in the margin causes far more errors than a large plain file does.
• Split enormous documents by purpose, not by size. If you only need the methodology and results sections, uploading those is better than uploading 400 pages and hoping attention lands in the right place.
• Remove personal data you do not need. Names, emails, and identifiers are rarely required for the analysis, and stripping them is faster than deciding whether you were permitted to upload them.
• Check that a PDF is actually text. Try selecting text in your PDF reader. If you cannot, it is a scan and OCR is involved, which changes how much you should trust extracted numbers.

Choosing the model for the file

This choice matters more than the prompt for document work.

• Gemini for anything long. It carries the largest context window of the models in Whizi, which means a 200 page report, a full contract set, or a dozen papers can be present at once rather than being processed in pieces. Cross-document questions only work properly when everything is genuinely in context.
• Claude for nuance. Legal language, policy documents, anything where the meaning depends on qualification and hedging. It is also the best at telling you what a document does not say, which is frequently the actual question.
• GPT for extraction. Invoices, forms, structured data, anything where you want a strict table or JSON back and consistent field names across a hundred rows.
• Any multimodal model for images, charts, screenshots, and diagrams. Ask it to describe what it sees before asking it to interpret, which catches misreadings early.

You can switch models mid-thread without re-uploading, which is the point of doing this in one workspace. Read with Gemini, extract with GPT, and have Claude write the summary, all against the same attached file. See switching models mid-conversation.

Working with several documents at once

Single document questions are useful. Multi document questions are where this stops being a convenience and starts being a capability you did not previously have.

Prompt: compare

Compare contract-a.pdf and contract-b.pdf. Return a table of every clause that differs: clause, what A says, what B says, and which is more favourable to us and why. Ignore differences in formatting and numbering. List clauses present in one and absent from the other separately.

Prompt: consolidate

Across all attached files, build one table answering [question]. One row per document, with the document name in the first column. Where a document does not address it, write NOT ADDRESSED. Do not merge documents that use different definitions without flagging it.

Prompt: find the contradiction

Where do these documents contradict each other? Quote both passages side by side and name the file each came from. Distinguish genuine contradictions from differences in scope or date.

Always reference files by name in your prompt once there are more than one. "In the vendor agreement, what is the notice period?" is unambiguous; "what is the notice period?" makes the model choose, and it will not tell you which it chose.

When the answer looks wrong

A short diagnostic sequence that resolves most problems.

The model says it cannot see the file. Confirm the attachment actually uploaded, then reference it explicitly: "In the attached report.pdf". For images especially, saying "look at the attached image" ensures the multimodal path is used rather than the model answering from the text of your question alone.

The answer is about the wrong document. Name the file in every question once the thread has several.

The numbers are subtly wrong on a scanned document. This is OCR, not the model. Ask it to quote the surrounding line verbatim so you can see what it read, then check the original.

It missed something you know is in there. Ask a locating question first: Quote the passage that discusses [topic] and give its page or section. If it cannot find it, the extraction is the problem, not the reasoning. Try a different model or upload the relevant pages separately.

It is summarising instead of answering. Ask for the evidence rather than the conclusion: Quote the exact text that supports your answer. This also catches the case where the answer was inferred from general knowledge rather than from your document, which is the most important failure to catch.

That last check deserves emphasis. A model asked about a well known contract type, a common regulation, or a famous paper can answer plausibly from training data without consulting your file at all. Requiring a quote forces the answer to be anchored to the document actually in front of it.

Privacy and what to think about before uploading

Files are stored in your workspace, are not used to train models, and can be deleted at any time. Each provider's data handling policy is available for review before you enable that model.

The judgment that remains yours is whether the document is one you are permitted to process with a third party service at all. The categories worth pausing on: personal data belonging to other people, anything covered by a specific confidentiality clause, material non public information, and regulated data such as health or financial records subject to particular handling rules.

The practical middle path for most people is to redact rather than abstain. Names replaced with roles, figures indexed rather than absolute, and identifiers removed will preserve everything the analysis needs while removing most of what makes a document sensitive.