3. Where an AI System Gets Its Answers From

A model generates. It does not consult.

That sentence is the whole reason retrieval exists, and it is worth sitting with, because it is the source of the behaviour that has done more than anything else to make enterprises nervous about this technology.

When a language model answers a question, it is producing text based on patterns absorbed during training. It has no index to check. It has no notion of a source. If the pattern it learned happens to be correct, the answer is correct. If the question sits in a region where the pattern is thin, the model will still produce a fluent, confident answer, because producing fluent confident text is what it does. This is what people call hallucination, and the word is misleading, since it implies a malfunction. The system is working exactly as designed.

Retrieval-augmented generation, usually shortened to RAG, is the architectural response. Before the model answers, a retrieval step searches a body of documents you control and pulls back the passages most relevant to the question. Those passages are handed to the model along with the question, and the model composes its answer from them.

The mechanics are less interesting than the consequence.

Two very different systems that look identical

Consider a system that answers questions about your compliance obligations. In one version, the model answers from training. In the other, the model retrieves the relevant clauses from your actual document set and answers from those.

To a user, these look the same. Both produce a paragraph of fluent, plausible prose. Both are quick. In a demonstration you would struggle to tell them apart.

They are not remotely the same object.

The first system has no source. When it is wrong, there is nothing to point at, no way to trace the error, and no way to correct it other than hoping the next model is better. The second system answered from a document you can open. When it is wrong, you can see which passage it used and why it went astray, and if the passage was out of date, you can fix the passage.

That difference is not a feature. It is a governance property, and it determines whether the system is auditable at all.

What it changes for you

The procurement question is simple, and few buyers ask it: where does the answer come from.

If a vendor's system answers from training, you are trusting the model's absorbed patterns and you have no recourse when they fail. If it answers from retrieval, you are trusting your own documents, and the failure modes become ones you can inspect and repair. The second is not automatically better in every case, but it is almost always better in a regulated or high-consequence setting, and it is the only one of the two that supports a serious audit conversation.

The follow-on questions are worth asking too. What corpus does the system retrieve from, and who controls it. How does it behave when the retrieval finds nothing relevant, since the honest response is to say so, and the tempting response is to answer anyway. Whether the retrieved passages are shown to the user or hidden behind the answer.

This connects directly to the traceability posture we hold across our products. A decision the system took, on the basis of a document you can open, approved by a person you can name, is a decision that will survive an audit. The retrieval architecture is a large part of what makes that possible.

What we are still working out

Long-context models complicate this. As models become able to hold very large documents in their working context, some of the work retrieval was invented to do can be done by handing the model the whole corpus and letting it find what it needs. This is more expensive and, at present, less precise, but the trend line is not obviously in retrieval's favour over a five-year horizon.

Our current view is that retrieval remains the right architecture for enterprise use even if long-context models make it technically optional, because retrieval gives you a traceable link between an answer and a source, and that link has governance value independent of its performance value. But we hold this position with some humility, and it is possible we are defending an architecture on grounds that a better one would also satisfy.

We are also still learning how much of the hallucination problem retrieval actually solves. It reduces it substantially. It does not eliminate it, because a model handed correct passages can still misread them. Anyone who tells you retrieval makes a system factual is overselling.

—

Mitochondria is an agentic AI product company based in Amsterdam and Pune. ISO 27001:2022 certified. Classified within the limited and minimal risk tiers of the EU AI Act, with controls aligned to the GDPR, UK GDPR and India's DPDP Act.