How Atlas grounds and cites answers

When you ask Atlas a question, the useful part is not merely the answer. You also need to see which project evidence supports it. Atlas grounds a response by retrieving material from the active project's sources and notes, drafting from that material, and attaching citations to important claims.

What this page covers

Learn how Atlas grounds answers, generates citations, marks evidence boundaries, and supports verification.

Grounding begins before chat

Atlas can retrieve evidence only after a source finishes processing. For a PDF, processing extracts text and associates passages with locations in the file. A website import captures the page's main content. A YouTube import depends on the available transcript.

If an import fails or omits content, chat cannot retrieve or cite the missing material. Check the source before relying on answers about it.

Retrieval narrows the context

Most projects contain more material than one answer can use. Atlas searches the active project for relevant passages, notes, source metadata, and other available context.

Question wording changes that search. A focused question gives Atlas a specific claim, method, source, or comparison to find. An @ mention narrows the search to the source, note, or artifact you selected.

Generation uses retrieved evidence

Atlas then drafts from the retrieved material. A grounded answer 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

A citation badge opens the source passage associated with a claim. Read the surrounding paragraph or page before using the answer in a note, report, or decision. The wider context shows whether the passage supports the wording and strength of the claim.

Why grounding can still fail

Grounding reduces unsupported answers, but it does not guarantee that every response is correct. 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."

Use grounded chat to locate and compare evidence. Keep the cited source open for the final judgment.

Verify an answer against its evidence

Atlas treats AI output as useful for research only when the reader can inspect where important claims came from. The goal is not to make AI sound confident. The goal is to make claims traceable.

The trust chain

The citation trust model has three parts:

1. A claim in an answer, summary, note, or generated artifact.
2. A source location that is supposed to support the claim.
3. A reader who checks whether the relationship is valid.

The third part matters. A citation is not proof by itself. It is an inspection path.

Why bare answers are not enough

Fluent answers can be wrong, incomplete, or overconfident. In research workflows, the cost of a wrong claim can be high: a literature review can misrepresent a paper, a business analysis can cite the wrong evidence, or a student can build an argument on a weak source.

That is why Atlas emphasizes cited answers, source jumps, and source-backed workflows.

What verification means

Verification means checking:

• whether the cited source is the right source
• whether the cited passage supports the exact claim
• whether nearby context changes the meaning
• whether the answer omits caveats
• whether another source disagrees
• whether the answer uses stronger language than the evidence allows.

The more important the output, the more carefully you should verify.

What Atlas should make easier

Atlas should reduce the cost of checking. The product should keep citations close to claims, make source jumps easy, preserve project context, and let users move between answer, source, note, and map without losing the evidence trail.

Refusal is part of trust

A trustworthy research system should sometimes say that the evidence is insufficient. Refusal or uncertainty is not a failure when the project genuinely lacks support for a claim.

Practical stance

Use AI output as a navigation and synthesis aid. Use the cited source as the authority. If the two disagree, trust the source and revise the AI output.

Understand citation behavior

Citations are Atlas's main verification mechanism. They connect an answer, summary, or generated artifact back to the source material that supports it.

What a citation means

A citation means Atlas found source evidence related to the claim. It does not mean the claim is automatically correct, complete, or strong enough for publication.

For important work, open the citation and inspect the passage yourself.

What to check

Citation states

Citation anchors

PDF citations may point to pages or extracted passages. Website and video citations depend on the imported content Atlas can access. If the source is scanned, transcriptless, blocked, or poorly extracted, citation quality will be lower.

How to improve citations

• Ask narrower questions.
• Mention the source, author, section, or concept.
• Ask Atlas to use only project sources.
• Ask for one cited claim per bullet.
• Open weak citations and ask Atlas to revise against the actual passage.