AI that cites its sources vs. AI that improvises

One name, two behaviors

Before answering, an AI assistant can look up your company’s documents and use that content to reply. The industry calls this technique RAG. On paper it sounds foolproof. In practice there are two ways to implement it that share the name and behave differently: the flexible one and the strict one.

The flexible version

It is the default setup. The system finds the document passages most similar to the question, hands them to the model, and the model uses them as a reference. The problem: it also feels free to fill in with its general knowledge whenever it thinks that helps.

That makes it useful for a marketing chat and dangerous for a regulated operation. It says true things with the same confidence it says invented ones. When a credit union advisor quotes a rate or a lawyer cites a ruling, the gap between a real fact and an invented one stops being a detail: it is a legal risk.

The strict version

It is the one we use when an answer has regulatory consequences. Three rules:

• Citation required. Every statement can be traced to the document that supports it. If there is no source, the agent answers “I could not find that information.”
• No filling in. The model does not add what it thinks it knows. If the documents do not cover the case, it hands it to a person.
• Reviewable record. Every answer is recorded along with the documents consulted, so anyone can check where it came from.

When to use each

The flexible version works for general guidance and first-line support, where “the closest match” is enough, mistakes are cheap and the user accepts confirming with a person.

The strict version is mandatory in tax, financial compliance, regulated products, health and any area where the client may have to defend an answer before an auditor. It is not an optional upgrade: it changes what the model is allowed to do.

What it takes

The strict version needs more preparation: well organized and labeled documents, ongoing quality reviews and an agent that can say “I don’t know” as naturally as it answers. The investment pays off the first time an answer goes through an audit without a single unsupported fact.

The rule of thumb: if your client may one day have to prove where an answer came from, choose the strict version from the start. Converting a flexible setup into a strict one once it is in production costs more than designing it right.