Skip to main content

Source Citations: Verified Data Values in Chat Answers

How the Source citations feature works — verification badges on values in chat answers, what Verified, Mismatch, Not found, and Unverifiable mean, and how to trace every number back to its source.

Written by Sabina Sokolowski

1. Overview

When an answer includes data from connected sources — portfolios, documents, market data — Source citations attaches a verification badge to each key value, showing where it came from. Unplex automatically re-checks each value against the original source data.

Verification is performed by the platform itself, not by the AI model: the displayed value is compared against the raw source output using deterministic rules, so an incorrect number cannot verify itself.

Note: Source citations is currently a beta feature. Where the setting described in Section 2 is not visible, a workspace administrator can request it be enabled.

2. Enabling Source Citations

  1. Open Settings (gear icon in the header) and go to Personalisation → Source citations.

  2. Switch on Validate source citations.

  3. Under Value types to cite, select which value types are marked: Numbers, Dates, Identifiers, Currency, Status, Names.

Note: marking more value types is more thorough but increases generation time for data-dense answers. Numbers and Currency are the recommended starting point where figures are the primary concern. This setting is personal and applies only to the configuring user's chats.

3. Verification Statuses

Status

Meaning

Verified

The displayed value matches the source data. Formatting differences (e.g. 1'250'000 vs 1,250,000), reasonable rounding, and % vs. decimal notation are recognized and accepted.

Mismatch

The displayed value does not match the source. The value should be treated with caution; the badge shows the difference.

Not Found

The value refers to a source reference that does not exist in the retrieved data — typically indicating the assistant cited something it did not actually retrieve.

Unverifiable

Primarily applies to longer text passages: the wording was rephrased enough that automatic comparison cannot confirm it. This is not an error; it indicates the passage requires manual judgment.

Only Verified values receive a positive indicator; every other status is flagged explicitly rather than accepted silently. Values are shown as highlighted chips during generation (a pulsing "Validating…" state) and resolve to one of the above statuses once the answer is complete.

4. Inspecting a Citation

Hovering over a badge shows a summary: field name, the tool or connector that produced the value, and the source value itself.

Clicking a badge opens the full citation card:

Source: the tool or connector the value came from (e.g. the portfolio system).

Field: the exact field in the source data (e.g. positions[0].market_value).

Source value vs. Displayed: the two values shown side by side, with the verification status.

Reason: shown for Mismatch or Unverifiable status, explaining what differs.

Original tool output: the complete raw data returned by the source, with the cited field highlighted — the underlying audit trail, allowing any value to be traced back to its exact source record.

5. Calculated Values

Where the assistant derives a figure rather than quoting one — a sum of positions, a percentage change, a difference between two values — the value appears as a Computed value badge. Opening it shows the expression used, the source variables, and the expected result (computed independently by the platform) alongside the displayed value.

The platform re-runs the calculation independently and compares the outcome with the assistant's stated value; a discrepancy in the assistant's arithmetic is shown as a mismatch without requiring the calculation to be checked manually.

6. Operational Notes

Automatic execution: verification runs at the end of every answer; no manual trigger is required.

Large data sets: where a source returns a very large number of values, coverage is capped and the assistant is informed explicitly that not every value can carry a citation, rather than presenting incomplete coverage as complete.

Unverifiable is not an error: a rephrased passage marked Unverifiable indicates the platform cannot machine-verify prose the way it verifies discrete numbers and dates.

Note: Source citations covers data values in chat answers, and is a distinct feature from the citation checker for legal references in documents (court decisions, statutes) — see How to Verify Citations & Compliance with Unplex.

Did this answer your question?