What research says about AI that cites its sources

"Plug a language model into your documents" — that's how most AI assistants are sold. Research is more demanding than that, and that is precisely where we built Manent.

What research says

The base technique is called RAG (retrieval-augmented generation): relevant passages are retrieved and given to the model before it answers (Lewis et al., 2020). But more recent work is clear: supplying documents does not guarantee a faithful answer. The model can still invent, blend two sources, or attribute a sentence to the wrong passage.

The safeguard that emerges from the literature is grounded generation with citation verification: the answer is broken down into claims, and each is verified against the sources before it is shown. Recent reviews of the field report that this discipline makes it possible to reach very high levels of citation accuracy — up to the order of 90 %+ — while sharply reducing hallucinations, where a "naive" RAG lets through many attribution errors.

What we do at Manent

We made this requirement a non-negotiable rule of the engine, not an option:

• Extractive answer first. When a source clearly states the answer, we cite it as is — no rewording that could go off the rails.
• Grounding verification. Every claim is checked against its sources (supported / unsupported / contradicted). An unsupported answer is not shown.
• Answering when it's sure. When a case falls outside what is documented, the assistant tells you and turns that moment into a card to be validated by your expert.
• Exact citation, always. File, section, page (or minute for a training video). The technician confirms on page 42 of their manual, in one second.

References

1. Lewis, P. et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS.
2. Systematic review of RAG (2020–2025), synthesis of more than a hundred works on grounding and citation verification (arXiv preprint, 2025).