You will be able to tell in advance when an uploaded file will be read well and when it will be misread or skipped.
When you upload a file to an assistant and ask about it, the answer can be wrong even though your question was clear. Most assistants now accept PDFs, word processing documents, spreadsheets and images, and many accept several files at once. Before the model can use any of them, the file has to be turned into text or tokens, because the model works on tokens (AI fundamentals lesson 3.1, "Your words become tokens before the model sees them"). Getting accurate answers from files starts with that conversion step, and it is where many file-based answers go wrong.
Digital PDFs and Word documents already contain their text. The assistant can pull it out directly, and this usually works well.
Spreadsheets are converted into rows and values. Sometimes the assistant runs code to read them. This is good for numbers, but formatting and notes can be lost on the way.
Images contain no text at all. This includes photos of printed pages and scanned PDFs. The text has to be recognised from the picture, which is text recognition. It is done either by a separate step or by a model that reads images directly, which AI fundamentals lesson 7.1, "How AI makes and reads images", explains.
You usually can't see which method was used, and the answer reads the same whether the conversion was perfect or full of errors.
Hui Min is about to rent a room in a condo in Toa Payoh (an invented example). The landlord's agent sends the tenancy agreement as phone photos of each printed page, 11 images in all, and asks her to sign by Friday. She uploads the images to an assistant and asks what the notice period is. The assistant says two months. When she reads the page herself, the clause says one month. The "1" was printed faintly and sits right next to a coffee stain, and it was read as a "2".
Her question was fine. The error happened when her photos were converted into something the model could read. She asks the agent for the original file, and he emails it within the hour.
Scanned and photographed pages like hers are the biggest risk. Faint print, shadows, skewed angles, stains and low resolution all cause misreadings, and a misread digit in a figure or a date looks just as confident as a correct one.
Complex tables come next. Merged cells, tables split across two pages and columns without clear borders can lose their structure, so a value ends up attached to the wrong row. Charts are harder still, because their numbers aren't written anywhere and have to be estimated from the shapes. Handwriting, including signatures, initials and margin notes, is often misread or ignored.
The rule to work from is that the further a document is from clean digital text, the more you should check.
A long document may not fit in the context window, the limit on how much material the model can work with at once (AI fundamentals lesson 5.1, "The context window is the model's only working memory"). Products deal with this in different ways. Some cut the file off at a certain length. Some summarise parts of it first. Some split it into pieces and search for the passages that look relevant to your question, which is retrieval (AI fundamentals lesson 5.3, "Retrieval: how an assistant looks things up").
With each of these approaches, some parts of the file may never be read when the model answers. Suppose you ask about a clause on page 40 of a 60-page contract. The answer might actually be based on pages 1 to 10, plus a guess. Even when the whole file fits, detail in the middle of long material is the most likely to be missed, as AI fundamentals lesson 5.2 explains.
File handling differs between products and changes often: which file types are accepted, the maximum file size, how many files you can add per chat, and whether long files are read in full or searched. Each product's help pages list these details. Read them for the assistant you use most, because they decide which of your documents will be handled well.
Two habits help whatever product you use. First, use the cleanest version of a document you can get. Ask for the digital PDF instead of a photo, and export a spreadsheet properly instead of taking a screenshot. Second, for long documents, upload only the part you need when you can, or ask about one section at a time. After that, check the answers, more carefully the further the source is from clean digital text, and especially where they involve figures, dates, tables, charts or handwriting.
You will see the difference between clean and messy files most clearly by testing both on the same questions. Find one clean digital PDF and one document you only have as a scan or a photo, and choose three questions whose answers you can check yourself in each. Ask the same questions about both documents and compare the answers.
Upload one clean digital PDF and one scanned or photographed document. Ask each the same three questions and compare how accurate the answers are.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).