Why your agent answers from your documents and not from the internet: retrieval, explained without jargon
When a customer asks your agent a question, two things happen in order: it searches your documents for the relevant bit, then it writes a reply from what it found. That ordering, search first and write second, is the whole reason a good agent quotes your price instead of inventing one. Here is how it works and how to write for it.
Quick answer
Retrieval means the agent searches your knowledge base for the passages most related to the customer's question and writes its reply only from those. It does not browse the internet or answer from general memory. If the search finds nothing relevant, the agent should say it is not sure. Retrieval works best on short, specific, standalone passages that use the words customers use, which is why price lists should be one line per item and FAQs should be written in the customer's phrasing.
The three steps
- Split. Your documents are cut into passages of a few sentences each when you add them.
- Search. The customer's question is matched against the passages, and the few best matches are pulled out.
- Write. The reply is composed from those passages and your instructions, and nothing else.
The model's job is language; your documents' job is facts. Keeping them separate is what makes the agent honest. The general definition is in what is a knowledge base.
When retrieval fails, and why
- The fact is spread across passages. The price on one page, the conditions on another. Fix: one line with both.
- The words do not match. Customer says "walk-in", document says "no appointment needed". Fix: include the customer's phrasing.
- The passage has no context. A table cell that says "RM 80" with the item name three rows up. Fix: item and price on one line.
- The document is a picture. A photo of a menu has no text. Fix: type it or export with text.
- Too much noise. Twelve blog posts drown two price pages. Fix: crawl the useful pages and remove the rest from the page list.
Writing for retrieval
Short standalone passages: one fact per line, subject named in the line, in plain words, in customer phrasing. "Scaling and polishing: RM 150 per visit, 30 minutes, adults only" retrieves perfectly. So does a FAQ entry that restates the question in the answer. The formats are in the price list guide and the FAQ guide. When the agent says it is not sure, it is usually a retrieval miss you can fix by writing the fact the way it was asked; see why the agent says it is not sure.
Frequently asked questions
Does the AI agent search the internet?
No. It searches only your knowledge base and answers from what it finds there.
Why does it sometimes miss a fact that is in my documents?
The fact was spread across passages, phrased differently from the question, in a table without context, or in an image. Rewrite it as one plain line.
Does retrieval mean my data trains the AI?
No. Retrieval reads your documents at answer time; nothing is trained on them.
How should I format documents for retrieval?
One fact per line, subject named, plain words, customer phrasing.
What happens when nothing matches?
The agent says it is not sure, hands over, and logs the question.
Write it the way they ask it
One line per fact, and the agent finds it every time. 7-day free trial.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.