You have a forty-page contract and you want three clauses pulled out of it. Or an Excel sheet from a client that needs summarising. These are among the jobs agents get handed most often.
In Orbi, uploading the document is enough. Text extraction runs on its own the moment the upload finishes, and the file's status moves from pending to done. After that you point the agent at the file by name.
Every uploaded file automatically gets a name, derived from the file name. If the name is already taken, a number is added to the end. So saying "review the report" needs no extra setup.
How each kind of document gets read
Each format is handled by its own logic, because flattening a table into plain text breaks it.
PDFs are read page by page, and if there is a text layer, the text comes from there.
In Excel files every sheet is converted to a markdown table. What gets read is the computed values, not the formulas themselves, so the agent gets the results.
Word documents are converted with their structure kept, headings stay headings. In presentations the text is taken in slide order.
For plain text and code files the character encoding is detected automatically. A Turkish file saved in an old Windows encoding also reads correctly, the letters do not break.
Images are not converted. They are viewed directly by a model that supports vision.
Scanned documents
A PDF may have no text layer. Scanned contracts and photocopies are like that. The pages are really images.
In that case the pages are converted to images and read by a vision model. The step handles at most thirty pages and works only with models from providers that have a key defined in your workspace.
The order is deliberate too. Normal text extraction is tried first, and image reading only kicks in when no text layer is found. The reason is cost, because reading an image is more expensive per page.
A three-hundred-page file does not fill the context
Putting a large document in front of the agent in one piece means filling the context and leaving no room for the actual work.
So reading is sliced. When an agent reads a file it takes a 12,000-character slice by default and moves forward as far as it needs. The slice size can be raised.
For "what does this contract say about termination", the agent does not need to read all forty pages. It finds the relevant place and carries on from there.
The inside of the file is searchable too
The text extracted from a document does not only reach the agent. It enters the workspace search as well.
So you can find a sentence from a report uploaded last month even if you do not remember the file name. Agents use the same search, which is why they can check "has this been discussed before" before starting a job.
Folders if you want them, none if you do not
The Files screen works without folders. If you never open one you see a flat list, and since search already looks at content, that is enough for most teams.
Those who want order can open folders, nest them and move files between them. The agents' tools are folder-blind, so you do not need to know where a file sits to name it.
Not imposing an empty folder structure is a deliberate choice. When you open a new workspace, you are not met by folders waiting to be filled.
Where a file came from
Agents produce files too. Look at a produced file and you see which task it came out of and which file it was derived from. Who uploaded it is written in the detail as well.
That small piece of information stops "what was this report based on" going unanswered a few weeks later.
Text documents and HTML files open inside the app, so you read them without downloading. You can also print from the same preview, which is the shortest route to seeing a report on paper before it goes to a client.
The only limit on file size is one hundred megabytes. Total storage depends on your plan.
Where to start
Pick the most tedious document you have. A long contract, or a report nobody wants to read.
Upload it, open a card and name the file. Do not make a summary your first request, because the accuracy of a summary is hard to measure. Ask a specific question instead and check the answer in the document yourself.
That first pass tells you whether the agent actually read the document. After that you can ask for a summary too.
Supported formats, page limits and the behaviour with scanned documents are covered in Files in the help center.
If you want to try it, you can open a free account and upload your first document.