Atlas · LLMs

Why grounded AI answers need citations

17 November 2026 · Burak Kara

Language models are fluent. That is exactly the problem: a fluent wrong answer looks the same as a fluent right one. In an organisation, where answers feed decisions, contracts and customer replies, “sounds right” is not good enough.

The fix is not a better prompt

You can ask a model to “only use the provided context” and to “say when it does not know”. Both help. Neither replaces a system that retrieves the right evidence first and shows where every claim came from.

A grounded setup has three parts:

  1. Retrieval. Search your own content (hybrid keyword and vector search tends to beat either alone) and hand the model only the relevant passages.
  2. Generation. The model writes an answer from those passages, not from memory.
  3. Citations. Each statement links back to the document it came from, so a person can verify it in seconds.

Permissions must be part of retrieval

If the search layer can return a document the user is not allowed to see, the model can leak it. Access control has to be applied inside the query, before anything reaches the model. Authorisation belongs to the platform, never to the model.

Measure it

Keep a set of real questions with known good answers. Re-run it whenever you change the model, the chunking or the prompt. Without that, every tweak is a guess.

This is the approach behind Atlas: connect your sources, retrieve with permissions applied, answer with citations.