{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"asking-your-own-documents","title":"What actually happens when you upload a file","url":"https://addaly.com/learn/ai-for-small-business/asking-your-own-documents","position":20,"of":88,"minutes":9,"answer":"Uploading a document may put all of it in front of the model or only the few passages your question retrieved, which is why pinpoint questions work well, exhaustive ones return confident incomplete lists.","keyIdea":"Uploading a document may put all of it in front of the model or only the few passages your question retrieved, which is why pinpoint questions work well, exhaustive ones return confident incomplete lists, and \"is this clause absent?\" is the least trustworthy question you can ask."},"body":{"format":"markdown","text":"## Not reading, in the way you mean it\n\nYou drag a 60-page supplier agreement into the chat and ask what the notice period is. An answer comes back in four seconds. It is easy to assume the tool read the document the way you would.\n\nTwo different things may have happened, and knowing which one changes how much you can trust the answer.\n\n**The whole document went in.** If the file fits inside the context window, its full text is placed in front of the model alongside your question. This is the better case. The model has seen every page.\n\n**Only some of it went in.** For larger files, or in tools built around document libraries, the file is chopped into passages: commonly a few hundred words each, with a little overlap so a sentence is not cut in half. Your question is used to find the handful of passages that look most relevant, and only those are given to the model.\n\nThe second approach is what makes it possible to search a thousand documents at once. It also means the model answered your question having seen perhaps four passages out of six hundred.\n\n## What this predicts\n\nThe behaviour that confuses people follows directly.\n\n**Pinpoint questions work well.** \"What is the notice period?\" The relevant passage is easy to find and the answer is in it. This is the strong case, and it is useful.\n\n**Questions needing the whole document fail quietly.** \"List every deadline in this contract.\" \"How many invoices mention overtime?\" \"Is anything in section 8 inconsistent with section 3?\" Retrieval fetched the passages that matched your words. It did not read the rest. You get a confident, incomplete list, and incompleteness is invisible: nothing is missing on the page in front of you.\n\n**Vague questions retrieve badly.** \"Is there anything I should worry about?\" matches nothing in particular, so the passages fetched are close to arbitrary.\n\n**Tables and figures often survive poorly.** A price table in a PDF becomes a run of numbers when the text is extracted. Ask about a figure in a table and check it against the original every time.\n\n> Figure: What an uploaded document will and will not tell you. Answers it gets right: What does clause 14.3 say; What is the notice period; Quote the clause about price rises; Explain clause 12 in plain English; What differs between these two versions. Answers it gets wrong quietly: List every deadline in this contract; Is there a non-compete clause; How many invoices mention overtime; Which of these forty quotes is dearest; Is anything in section 8 inconsistent. Retrieval fetched the passages that matched your words and did not read the rest, so an incomplete list looks exactly like a complete one. There is no clause about Y is the least trustworthy sentence in this whole area.\n\n## Asking better\n\nFour habits, all cheap:\n\n**Ask about one thing at a time.** Six separate questions beat one question with six parts, because each gets its own retrieval.\n\n**Use the document's own words.** If the contract says \"termination for convenience\", search with that phrase rather than \"cancelling early\". Retrieval matches on meaning rather than exact words, which helps, but the document's vocabulary still retrieves better.\n\n**Ask for the quote and the page.** \"Quote the exact clause and give the section number.\" This turns an answer into something you can verify in ten seconds, and it exposes the case where the model produced a plausible clause that is not there.\n\n**Go section by section for anything exhaustive.** If you need every deadline, paste section 4 and ask, then section 5. Tedious, and it is the only way to get a complete list.\n\n## The honest limits\n\nThree things a document tool will not do reliably, whatever the marketing says.\n\nIt will not tell you that something is **absent**. \"Does this contract contain a non-compete?\" is a hard question, because failing to retrieve a clause looks exactly like the clause not existing. A confident \"no non-compete clause is present\" is the least trustworthy sentence in this whole area.\n\nIt will not **compare across many documents** by counting. \"Which of these forty quotes had the highest labour rate?\" is a spreadsheet question.\n\nIt will not replace **reading the thing that matters**. A lease, a loan, a supply agreement you will live with for three years: read it. Use the tool to find the clauses, explain the language and generate the questions. Then read it yourself, and if the money is significant, have somebody qualified read it too.\n\n## A reasonable working rule\n\nTreat an uploaded document as a very fast, slightly careless assistant who will find you the right page nine times out of ten and will never tell you when it failed. That is valuable, and it is not the same as having read the document.\n\n## Before you upload anything\n\nTwo checks, ten seconds each, that prevent most of the disappointment.\n\n**Is it text, or a picture of text?** A PDF produced by a computer contains selectable text. A PDF made by photographing pages contains images. If you cannot select a sentence with your cursor, neither can the tool, and it will need optical character recognition first, which is a later lesson in this block and comes with its own error rate.\n\n**Does it need to leave your building?** A supplier's price list is one thing. Employee contracts, medical notes, a client's confidential brief are another. The upload button does not ask you this question, so you have to.\n\n## The habit worth forming\n\nAsk your question, then ask a second one: *quote the exact text this came from.* Read the quote against the document. Ten seconds.\n\nDo this on every document answer for the first fortnight. You will learn very quickly which kinds of question your particular tool handles well and which it fudges, and that knowledge transfers to every document you upload afterwards."},"check":{"question":"An owner uploads twelve supplier contracts and asks which ones contain a price-escalation clause. She gets a clear list of three. What is the main reason to distrust it?","options":[{"text":"Retrieval fetched the passages that matched her wording, so contracts phrasing the same clause differently were never looked at and their absence from the list means nothing.","correct":true},{"text":"Twelve documents is more than most tools can process at once, so some files were probably skipped entirely during upload.","correct":false,"tempting":"Blames an upload limit that most tools do not have at this size, and the same failure occurs with three documents that are all successfully indexed."},{"text":"Contract language is legally technical, so the model may have misclassified a clause it did retrieve and read correctly in full.","correct":false,"tempting":"A genuine risk of misreading, but it would produce a wrong entry on the list rather than explaining why the list may be missing several contracts altogether."},{"text":"The model cannot compare documents against each other, so any question spanning more than one file will produce an unreliable answer regardless of how carefully the question is phrased or how many files are involved.","correct":false,"tempting":"Overstates the limit into a blanket rule; comparing a small number of retrieved passages is something these tools do adequately when the passages were actually found."}]},"prev":{"slug":"testing-a-prompt","title":"Ten cases before you trust it","url":"https://addaly.com/learn/ai-for-small-business/testing-a-prompt"},"next":{"slug":"why-retrieval-beats-memory","title":"Grounding: answering from your material rather than its memory","url":"https://addaly.com/learn/ai-for-small-business/why-retrieval-beats-memory"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What actually happens when you upload a file', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/asking-your-own-documents (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What actually happens when you upload a file","courseTitle":"AI for a Small Business","module":{"index":3,"slug":"your-own-knowledge","title":"Your own documents, and getting straight answers out of them"},"answer":null,"keyIdea":"Uploading a document may put all of it in front of the model or only the few passages your question retrieved, which is why pinpoint questions work well, exhaustive ones return confident incomplete lists, and “is this clause absent?” is the least trustworthy question you can ask.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 3.1","figure":{"left":{"title":"Answers it gets right","items":["What does clause 14.3 say","What is the notice period","Quote the clause about price rises","Explain clause 12 in plain English","What differs between these two versions"]},"type":"compare","right":{"title":"Answers it gets wrong quietly","items":["List every deadline in this contract","Is there a non-compete clause","How many invoices mention overtime","Which of these forty quotes is dearest","Is anything in section 8 inconsistent"]},"title":"What an uploaded document will and will not tell you","caption":"Retrieval fetched the passages that matched your words and did not read the rest, so an incomplete list looks exactly like a complete one. There is no clause about Y is the least trustworthy sentence in this whole area."},"alt":"Figure: What an uploaded document will and will not tell you. Answers it gets right: What does clause 14.3 say; What is the notice period; Quote the clause about price rises; Explain clause 12 in plain English; What differs between these two versions. Answers it gets wrong quietly: List every deadline in this contract; Is there a non-compete clause; How many invoices mention overtime; Which of these forty quotes is dearest; Is anything in section 8 inconsistent. Retrieval fetched the passages that matched your words and did not read the rest, so an incomplete list looks exactly like a complete one. There is no clause about Y is the least trustworthy sentence in this whole area."}},"blocks":[{"t":"h2","id":"not-reading-in-the-way-you-mean-it","runs":[{"text":"Not reading, in the way you mean it"}]},{"t":"p","runs":[{"text":"You drag a 60-page supplier agreement into the chat and ask what the notice period is. An answer comes back in four seconds. It is easy to assume the tool read the document the way you would."}]},{"t":"p","runs":[{"text":"Two different things may have happened, and knowing which one changes how much you can trust the answer."}]},{"t":"p","runs":[{"text":"The whole document went in.","b":true},{"text":" If the file fits inside the context window, its full text is placed in front of the model alongside your question. This is the better case. The model has seen every page."}]},{"t":"p","runs":[{"text":"Only some of it went in.","b":true},{"text":" For larger files, or in tools built around document libraries, the file is chopped into passages: commonly a few hundred words each, with a little overlap so a sentence is not cut in half. Your question is used to find the handful of passages that look most relevant, and only those are given to the model."}]},{"t":"p","runs":[{"text":"The second approach is what makes it possible to search a thousand documents at once. It also means the model answered your question having seen perhaps four passages out of six hundred."}]},{"t":"h2","id":"what-this-predicts","runs":[{"text":"What this predicts"}]},{"t":"p","runs":[{"text":"The behaviour that confuses people follows directly."}]},{"t":"p","runs":[{"text":"Pinpoint questions work well.","b":true},{"text":" “What is the notice period?” The relevant passage is easy to find and the answer is in it. This is the strong case, and it is useful."}]},{"t":"p","runs":[{"text":"Questions needing the whole document fail quietly.","b":true},{"text":" “List every deadline in this contract.” “How many invoices mention overtime?” “Is anything in section 8 inconsistent with section 3?” Retrieval fetched the passages that matched your words. It did not read the rest. You get a confident, incomplete list, and incompleteness is invisible: nothing is missing on the page in front of you."}]},{"t":"p","runs":[{"text":"Vague questions retrieve badly.","b":true},{"text":" “Is there anything I should worry about?” matches nothing in particular, so the passages fetched are close to arbitrary."}]},{"t":"p","runs":[{"text":"Tables and figures often survive poorly.","b":true},{"text":" A price table in a PDF becomes a run of numbers when the text is extracted. Ask about a figure in a table and check it against the original every time."}]},{"t":"h2","id":"asking-better","runs":[{"text":"Asking better"}]},{"t":"p","runs":[{"text":"Four habits, all cheap:"}]},{"t":"p","runs":[{"text":"Ask about one thing at a time.","b":true},{"text":" Six separate questions beat one question with six parts, because each gets its own retrieval."}]},{"t":"p","runs":[{"text":"Use the document’s own words.","b":true},{"text":" If the contract says “termination for convenience”, search with that phrase rather than “cancelling early”. Retrieval matches on meaning rather than exact words, which helps, but the document’s vocabulary still retrieves better."}]},{"t":"p","runs":[{"text":"Ask for the quote and the page.","b":true},{"text":" “Quote the exact clause and give the section number.” This turns an answer into something you can verify in ten seconds, and it exposes the case where the model produced a plausible clause that is not there."}]},{"t":"p","runs":[{"text":"Go section by section for anything exhaustive.","b":true},{"text":" If you need every deadline, paste section 4 and ask, then section 5. Tedious, and it is the only way to get a complete list."}]},{"t":"h2","id":"the-honest-limits","runs":[{"text":"The honest limits"}],"role":"limit"},{"t":"p","runs":[{"text":"Three things a document tool will not do reliably, whatever the marketing says."}]},{"t":"p","runs":[{"text":"It will not tell you that something is "},{"text":"absent","b":true},{"text":". “Does this contract contain a non-compete?” is a hard question, because failing to retrieve a clause looks exactly like the clause not existing. A confident “no non-compete clause is present” is the least trustworthy sentence in this whole area."}]},{"t":"p","runs":[{"text":"It will not "},{"text":"compare across many documents","b":true},{"text":" by counting. “Which of these forty quotes had the highest labour rate?” is a spreadsheet question."}]},{"t":"p","runs":[{"text":"It will not replace "},{"text":"reading the thing that matters","b":true},{"text":". A lease, a loan, a supply agreement you will live with for three years: read it. Use the tool to find the clauses, explain the language and generate the questions. Then read it yourself, and if the money is significant, have somebody qualified read it too."}]},{"t":"h2","id":"a-reasonable-working-rule","runs":[{"text":"A reasonable working rule"}]},{"t":"p","runs":[{"text":"Treat an uploaded document as a very fast, slightly careless assistant who will find you the right page nine times out of ten and will never tell you when it failed. That is valuable, and it is not the same as having read the document."}]},{"t":"h2","id":"before-you-upload-anything","runs":[{"text":"Before you upload anything"}]},{"t":"p","runs":[{"text":"Two checks, ten seconds each, that prevent most of the disappointment."}]},{"t":"p","runs":[{"text":"Is it text, or a picture of text?","b":true},{"text":" A PDF produced by a computer contains selectable text. A PDF made by photographing pages contains images. If you cannot select a sentence with your cursor, neither can the tool, and it will need optical character recognition first, which is a later lesson in this block and comes with its own error rate."}]},{"t":"p","runs":[{"text":"Does it need to leave your building?","b":true},{"text":" A supplier’s price list is one thing. Employee contracts, medical notes, a client’s confidential brief are another. The upload button does not ask you this question, so you have to."}]},{"t":"h2","id":"the-habit-worth-forming","runs":[{"text":"The habit worth forming"}]},{"t":"p","runs":[{"text":"Ask your question, then ask a second one: "},{"text":"quote the exact text this came from.","i":true},{"text":" Read the quote against the document. Ten seconds."}]},{"t":"p","runs":[{"text":"Do this on every document answer for the first fortnight. You will learn very quickly which kinds of question your particular tool handles well and which it fudges, and that knowledge transfers to every document you upload afterwards."}]}],"check":{"question":[{"text":"An owner uploads twelve supplier contracts and asks which ones contain a price-escalation clause. She gets a clear list of three. What is the main reason to distrust it?"}],"options":[{"text":[{"text":"Retrieval fetched the passages that matched her wording, so contracts phrasing the same clause differently were never looked at and their absence from the list means nothing."}],"correct":true,"tempting":null},{"text":[{"text":"Twelve documents is more than most tools can process at once, so some files were probably skipped entirely during upload."}],"correct":false,"tempting":[{"text":"Blames an upload limit that most tools do not have at this size, and the same failure occurs with three documents that are all successfully indexed."}]},{"text":[{"text":"Contract language is legally technical, so the model may have misclassified a clause it did retrieve and read correctly in full."}],"correct":false,"tempting":[{"text":"A genuine risk of misreading, but it would produce a wrong entry on the list rather than explaining why the list may be missing several contracts altogether."}]},{"text":[{"text":"The model cannot compare documents against each other, so any question spanning more than one file will produce an unreliable answer regardless of how carefully the question is phrased or how many files are involved."}],"correct":false,"tempting":[{"text":"Overstates the limit into a blanket rule; comparing a small number of retrieved passages is something these tools do adequately when the passages were actually found."}]}],"order":[2,3,0,1],"why":null},"prev":{"slug":"testing-a-prompt","title":"Ten cases before you trust it","minutes":8},"next":{"slug":"why-retrieval-beats-memory","title":"Grounding: answering from your material rather than its memory","minutes":9}}}