What it is, without the jargon
A language model knows how to talk, but it knows nothing about your business. RAG means that, before answering, the agent searches your documents for the chunks that matter and writes the answer with them in front of it. It doesn’t answer from memory: it answers by reading.
How it searches
- 1It reads the question in context. It doesn’t search with the last message alone: it also uses the customer’s earlier messages, so “what if I already opened it?” is understood as a question about returns.
- 2It searches two ways at once. By meaning, to find what’s said in other words, and by exact words, so it doesn’t miss an order number or a reference. Both lists are merged into one.
- 3It answers with the sources in front of it. The chunks it found go into the answer with their name and page. If what it got isn’t enough, the agent can search again.
Why it doesn’t make up a number
The agent has explicit rules: it doesn’t state a deadline, a price, a condition or a guarantee it hasn’t read in what it found. It doesn’t make up IDs, amounts or dates.
If the search finds nothing, it says it isn’t in the documentation and offers to hand off to a person. That beats an answer that sounds right and isn’t true.
Documents for the general, systems for the specific
A document tells you how many days you have to return an order. It doesn’t tell you where this customer’s order is. For anything tied to a specific account — an order, a charge, a booking — the agent doesn’t search documents: it queries the system with the actions you turned on. And if no action covers it, it says so and hands it to your team.
You can see what it read
Every answer leaves in its trace what was searched, how many chunks came up and which sources they came from, with their match score. If an answer doesn’t convince you, you can see where it came from.
