How Referencing & Hallucination Checks Work
6 min read
Overview
When Unique AI Conduct uses information that comes from outside its own reasoning — your knowledge base, the public web, a file it opens, or a connected tool — it does two things to help you trust the result. First, it references every fact back to its source. Second, for most of those sources, it runs an automatic hallucination check to confirm the answer is actually supported by what it retrieved.
This page explains how both mechanisms work, what you see in the chat, and when the hallucination check does and does not apply. It expands on the Referencing & Citations and Hallucination Check capabilities described on the main Unique AI Conduct page.
Referencing happens automatically whenever the agent retrieves information — you never have to ask for it. The hallucination check also runs on its own, right after the agent finishes its answer.
Referencing & Citations
Where references come from
Conduct adds references in four situations:
Scenario | When it happens | What the reference links to |
|---|---|---|
Knowledge base search | The agent searches your organization's uploaded documents to find relevant information. | The exact document (and page) the fact came from. |
Web search | The agent searches the public web, within your organization's egress policy. | The external web page, shown with its title and link. |
Reading uploaded or opened files | The agent opens a file you attached (or one already in the conversation) and uses facts from it. | The file, and the specific page that was read. Page-level detail is available for PDF and Word files; other formats are referenced at document level. |
Connected tools (MCP) | The agent calls an external connected tool or integration. | The tool's result, shown as a clickable link with a readable title. |
How a reference is created
Behind the scenes, referencing is a hand-off between the agent and the platform:
The agent retrieves and records the source. Each retrieval action — a knowledge base search, a web search, a request to cite an opened file, or a connected-tool call — writes the source it used into a small per-turn record called a manifest, and labels each result with a short citation marker. For example,
[source1]identifies a knowledge base result,[websource1]a web result,[filesource1]a page of an opened file, and[mcpsource1]a connected-tool result.The agent cites as it writes. When the agent states a fact that came from a source, it places the matching marker immediately after it — for example, "Quarterly revenue rose 15% [source1]."
The platform turns markers into references. Once the agent finishes, the platform reads the manifest, replaces each marker with a numbered footnote, and attaches a clickable reference chip that opens the original document, web page, or file page.
For knowledge base results, the manifest stores the actual text snippets — called chunks — that the search returned, not just a link. This detail matters for the hallucination check described below.
What you see in the chat
Small numbered footnotes (1, 2, …) next to the facts in the answer.
Clickable reference chips that take you straight to the source document, web page, or file page.
Numbering is continuous within a single answer and starts fresh on each new turn.
Footnote numbers are grouped by source type — knowledge base first, then web, file, and connected tools — rather than by where the fact appears in the answer. A fact mentioned earlier can therefore carry a higher number than one mentioned later.
References and the hallucination check apply only to the chat message the agent writes back to you. They do not reach into files the agent generates (such as an Excel report or a chart): those are delivered as downloadable deliverables and are not citation-linked or checked. Likewise, if the agent answers purely from its own general knowledge, there is nothing to cite — and nothing to verify.
Hallucination Check
What it does
The hallucination check is an automatic review that runs after the agent produces its answer. It compares the claims in the answer against the source material the agent actually retrieved, and flags statements that are not supported by those sources. The outcome is shown as an indicator on the message, so you can see at a glance how well-grounded the answer is.
How it works
While answering, the agent retrieves sources. These are stored — together with their text — in the per-turn manifest.
After the answer is complete, the platform gathers the retrieved source text: for the knowledge base, the retrieved chunks; for the web, the content of the pages the agent read; for a cited file, the page is independently re-derived as text, or as a rendered image where the agent read the page visually.
A dedicated evaluation compares the answer to that source text and identifies any claims the sources do not back up.
The result is surfaced on the message as a hallucination indicator.
When the hallucination check runs
Scenario | Hallucination check |
|---|---|
Knowledge base search | Checked |
Web search | Checked |
Reading uploaded or opened files | Checked The platform independently re-derives the cited page — as text, or as a rendered image for pages the agent read visually — and compares it to the claim. |
Connected tools (MCP) | Partially checked MCP results are referenced and do factor into the check, but at lower fidelity than the other sources: the cited text is captured in full only when the connector's response included it, and otherwise only a truncated summary reaches the check. Treat MCP-sourced claims with extra caution. |
Generated files | Not checked |
The hallucination check covers knowledge base results, web results, and facts cited from files the agent opens. Results retrieved through connected tools (MCP integrations) are covered only partially, and content inside files the agent generates is not covered at all — generated-file content is neither referenced nor checked. Always review those claims yourself.
Limitations
Connected tools (MCP) are only partially checked. Information the agent retrieves through connected tools is referenced — answers include clickable source links with readable titles — and it does reach the hallucination check, but with lower fidelity than knowledge base, web, or file sources. Where the connector's response did not include the underlying text, only a truncated summary is available to the check. Confirm those results yourself.
Generated files are not covered. Referencing and the hallucination check apply only to the references added to the chat message the agent writes back to you. Content inside files the agent generates (for example, an Excel report or a chart) is delivered as a downloadable file and is neither citation-linked nor hallucination-checked.
Page-level references depend on the file format. References to an opened file point at the exact page for PDF and Word documents. PowerPoint, Excel, HTML, and other formats are referenced at document level.
Referencing depends on the agent using retrieval. Reference chips appear only for information the agent retrieves through knowledge base search, web search, file citation, or a connected tool. Claims drawn from the agent's general knowledge are not cited and cannot be checked.
The check verifies grounding, not source accuracy. The hallucination check confirms that the answer matches the retrieved sources. It does not judge whether the underlying source is itself correct or up to date.