Giving a Model Your Own Data

Why a model doesn't know your data

Its knowledge stops at a date, and never included your files

Scroll

Its knowledge stops at a date, and never included your files

A model knows what was in its training data and nothing else. That data was collected up to a cutoff date, and it was mostly public: web pages, books, code. Your company's policies, last week's tickets and the contract in your inbox were never in it.

Ask about any of those and the model can do one of two things. It can say it doesn't know, or it can write a plausible answer from what similar documents usually say. The second happens more often than you would like, and it reads exactly like a correct answer.

The fix is to put the facts in the prompt. For one document you can paste it in. For thousands you can't: they will not fit in the context window, and sending all of them with every question would be slow and expensive. The job becomes finding the few passages that matter for this question and sending only those. That is retrieval, and the pages on chunking, vector indexes and search build it piece by piece.

Retrieval
IT ONLY KNOWS WHAT IT WAS TRAINED ON What is our refund window? model Usually 30 days. a guess What is our refund window? + refund-policy.pdf model 45 days, per section 3.2. grounded but you cannot attach everything, every time: context window all your files One attached document grounds the answer. Thousands will not fit, so something has to choose which passages to send. That is retrieval.
The same question with and without the document, and why you cannot just attach everything.