Skip to main content

How Atlas grounds answers in sources

Atlas is designed for research work where an answer is only useful if you can inspect the evidence behind it. Grounding is the process of making an answer depend on the sources, notes, and context inside the active project instead of relying only on general model knowledge.

Grounding begins before chat

When you import a source, Atlas has to process it before it can be used well. For a PDF, that may include reading extracted text and linking passages back to document locations. For a website, it means capturing the main content. For a video, it depends on transcript availability.

If processing is incomplete, the later answer will be incomplete too. Chat cannot cite evidence that never entered the project.

Retrieval narrows the context

Most projects contain more material than can fit into a single answer context. When you ask a question, Atlas retrieves the pieces that appear relevant: passages, notes, source metadata, or other project context.

This is why wording matters. A focused question gives retrieval a clearer target. Mentions make the target even clearer because they tell Atlas which source, note, or artifact should matter.

Generation uses retrieved evidence

After retrieval, Atlas drafts an answer from the available evidence. Good grounded answers should:

  • answer the question asked
  • use the retrieved project context
  • cite important claims
  • avoid stronger claims than the sources support
  • acknowledge when evidence is missing or conflicting.

Citations make the answer inspectable

Citations are the bridge between generated text and source evidence. They let you move from an answer back to the material Atlas used.

That bridge is valuable because it keeps the reader in control. You can open the source, read the surrounding paragraph, and decide whether the answer is faithful.

Why grounding can still fail

Grounding reduces unsupported answers, but it is not magic. Problems can still happen when:

  • the source was not imported
  • the source failed processing
  • the question is too broad
  • retrieval found the wrong passage
  • the source itself is ambiguous
  • the answer overstates weak evidence
  • citations point to related but insufficient context.

When that happens, narrow the question, name the source, ask for cited evidence, or inspect the source manually.

How to get better grounded answers

The strongest prompts are specific about evidence and scope:

  • "Using only the uploaded paper, what evidence supports X?"
  • "Compare Source A and Source B on this claim."
  • "Cite each bullet."
  • "If the sources do not answer this, say so."

Grounded chat is most reliable when you treat it as an evidence-navigation system, not just a text generator.