Help for owners
My AI assistant answers from my company documents, but gets it wrong
Short answer
When an assistant answers from your documents, it first searches them for passages and only then writes a reply. Most wrong answers come from the search step: the wrong passage, a split table, an old file, or no passage at all. Pause what it may say, collect bad examples, and check the sources it used.
Does this sound like your situation?
- It quotes a price, policy or date that is not in my documents.
- It answers from an old version of a file has already replaced.
- Questions about a table or a PDF scan get nonsense answers.
- It answers one question well and a nearly identical one badly.
- It never says it does not know, it always sounds sure.
Why does this happen?
The search step picks the wrong passages
Documents are cut into pieces and stored so they can be searched by meaning. OpenAI's documentation describes this: files are split into chunks (800 tokens with 400 tokens of overlap by default), and a search returns the most similar chunks with a similarity score. If the right chunk does not rank high enough, the model answers from whatever came back instead.
Confirmed by developers.openai.com
The cut falls in the wrong place
A price list, a table or a policy with exceptions can be split so that a rule ends up in one piece and its exception in another. The assistant then sees half of the answer and states it with full confidence.
Old and new versions both sit in the index
If last year's price sheet and this year's are both uploaded, the search can return either one. Nobody told the system which file is the current one.
Scans and layouts are not read as text
A scanned PDF, a photo of a page or a multi-column layout can come out as scrambled text, so the passage that gets stored does not say what the page says.
Nothing tells the assistant to stay quiet
Unless it is set up to answer only from the passages found, and to say so when none fit, the model fills the gap from its general knowledge. That is where invented answers come from.
What can I do about it today?
- Stop it from answering sensitive questionsIf it talks about prices, refunds, health, legal or safety matters, switch those topics off or add a visible line that a person confirms them. Do not delete anything.
- Keep a list of bad answersSave the question, the answer and what the right answer should be, with a screenshot. Twenty real examples are worth more than any general complaint.
- Check the documents themselvesFor each bad answer, open the file the answer should have come from. Is it the current version? Is the text selectable, or is it a scan? Is the answer actually written down?
- Remove or rename outdated filesMove old versions out of the folder the assistant reads, or add the date to the file name. Keep the originals somewhere safe first.
- Ask the vendor or builder which passages it usedA good setup can show the passages behind each answer. If yours cannot, that is the first thing to fix, because without it nobody can tell a search error from a writing error.
When should I bring someone in?
- Customers or staff already act on its wrong answers.
- The documents are scans, tables or many languages and the answers stay bad after cleaning them.
- Nobody can show which passage an answer came from.
How do I handle it?
- Build a test set from the bad answersI turn your saved examples into a list of questions with the right answers, so every change can be measured instead of guessed.
- Separate search errors from writing errorsFor each failure I look at the passages that were retrieved. If the right one is missing, the fix is in the documents or the search. If it was there and ignored, the fix is in the instructions.
- Fix the documents and the cuttingText extraction, chunk size and overlap, metadata such as date and version, and removal of duplicates are adjusted, then the test set is run again.
- Make it able to say it does not knowThe assistant is limited to the passages found, shows its source and hands over to a person when nothing fits.
- Re-run the test set on every changeNew documents or a new model can break what worked, so the same questions are checked again before release.
The price depends on what I find, so I quote in writing after a short discovery call. Tell me what is happening and I reply within 24 hours.
What do owners ask about this?
Do I need a better AI model?
Usually not first. Wrong answers over your own files most often come from the search step or the documents. A stronger model reading the wrong passage still gives a wrong answer.
Can I just upload more documents?
More files can make it worse if they contradict each other or repeat old versions. Fewer, current, clean documents often beat a big pile.
Will it ever be 100 percent right?
No system built on language models can promise that. The goal is to measure how often it is wrong, show sources, and send doubtful questions to a person.
Is this the same as an AI feature that invents answers in general?
It overlaps. This page is about assistants that search your documents. The wider page on wrong answers is linked below as a related help page.
Sources
How was this page made?
I drafted this page with AI assistance from the official documents listed under Sources. Found a mistake? Tell me.