Evidence and citations

What are Evidence and citations?

In DataFact, Evidence is supporting material returned with analytical work when the Agent and workflow provide it. A citation is a locator that links part of an answer to source content returned by retrieval. Depending on the source and task, supporting material can also include query context, rows, charts, generated files, or visible execution information.

A citation makes a source easier to inspect; it does not automatically validate the answer. The retrieved passage may mention the topic without supporting the conclusion, omit an important qualifier, or come from a stale or non-authoritative document. A database result may need a reproducible query and an approved metric definition rather than a document citation.

Evidence is conditional. A database analysis, model-only answer, tool result, or unsupported source can produce an answer without a document citation. A citation can also fail to open when the source is unavailable, indexing is incomplete, access has changed, or the returned locator cannot be resolved. Treat the artifacts that are present as material for review, not as proof by themselves.

Why it matters

Important answers often need to be checked by someone who did not ask the original question. A useful locator reduces the effort required to find the relevant source and read it in context.

How DataFact uses Evidence

When a compatible workflow returns a citation, DataFact renders the locator with the answer and can open the corresponding source context. Other workflows may return different analytical artifacts or no citation.

Example

If an answer attributes a margin change to a surcharge in a logistics review, open the cited passage and confirm the route, date, and exception. Then reproduce the database number separately if the conclusion also depends on structured data.

Verification questions

• Does the source support the specific claim, not merely mention the topic?
• Is the source current and authoritative for this decision?
• Are important qualifiers omitted from the answer?
• For numbers, can the result be reproduced from the underlying data?

Limitations and edge cases

• A citation can be relevant to the subject without supporting the specific statement.
• Source permissions or retention can make an earlier citation unavailable.
• A database result can be inspectable without having a document citation.
• A citation does not expose complete claim-level provenance for every sentence.

Related concepts

• AI answer verification
• How an answer is produced
• Knowledge and answer quality
• Evaluation lifecycle