What the documents look like
Creator-hosted sales campaigns centre on external creator fulfilment and commission settlement. Brand commerce staff, MCN agencies and creator managers work across four document types: creator cooperation agreements, commission setting sheets, sample shipment sheets and livestream performance reports. Each campaign produces a batch of transaction details, and after-sales commission deduction checks often span weeks.
Citation tracing for this category answers a narrow question: can a statement made during order fulfilment or after-sales support be matched to a specific document and a specific document state. The four document types carry different fields. A cooperation agreement records agreed commission percentages and signing dates. A commission setting sheet records the commission ratio, an effective date and a review status. A sample shipment sheet records a shipment number, item quantity and receipt status. A livestream performance report records per-session transaction figures. A conclusion drawn in an after-sales reply may combine fields from more than one of these documents, and the tracing step must keep each field attached to its own source.
Constraints at the citation tracing step
The knowledge base search node runs retrieval against the selected knowledge bases and outputs cited content as an array. The array length may be zero, and the output path still completes when nothing is retrieved. This output is handed to the answering stage, where each conclusion should correspond line by line to cited source text.
The minimum relevance value is a filter threshold. It applies only during semantic retrieval or when result reranking is used. It does not perform permission filtering or entity isolation, and it does not exclude unauthorised records or unrelated businesses.
Hybrid retrieval combines multiple retrieval paths through RRF. RRF is a fusion process inside hybrid retrieval, and it is not a separate toggle. Result reranking requires a rerank model and must be enabled in the settings; when it is not enabled, that path does not run. Image retrieval requires an image input and a model with the corresponding capability, so it does not execute on every request.
Knowledge bases that use different vector models cannot be associated together. The vector model identifier of each base is checked first, and bases that do not match are excluded.
Hypothetical paper exercise. Source text one: "Creator A and Brand B signed a creator cooperation agreement on 1 March 2026, agreeing that the campaign commission is settled at fifteen percent of transaction value." Source text two: "Sample shipment sheet number Y-0315, three sample items shipped, receipt status signed for." Source text three: "The commission setting sheet shows a commission ratio of fifteen percent, an effective date of 5 March 2026, and a review status of pending review." Draft sentence to be assessed: "Creator A's campaign commission has been settled at fifteen percent."
The assessment requires the same subject, the same period and the same field. The agreement ratio and the setting sheet ratio match, but the setting sheet review status is pending review. A pending review does not constitute approval, and it does not constitute settlement execution.
Configuration and implementation checks
| Configuration or check item | Use on this page | Correspondence with the sample |
|---|---|---|
| Associated knowledge bases | Select the bases holding the cooperation agreement, commission setting sheet, sample shipment sheet and livestream report | Check each base's vector model identifier first; associate only matching bases and exclude the rest |
| Citation limit | Controls the number of tokens cited per search | For this exercise the starting value is set to cover the three source texts; compare it against what retrieval actually returns before adjusting |
| Minimum relevance | Filter threshold during semantic retrieval or result reranking | Acts as a filter only; it is not used to exclude unauthorised records or unrelated subjects |
| Result reranking | Second-pass ordering of candidate text | Requires a rerank model enabled in the settings; when it is not enabled this path does not run |
| Cited content output | Receives the citation array returned by retrieval | The array length may be zero; when it is empty, check how the downstream answering stage is designed to handle it |
| Manual comparison step | Compare each answer statement against the cited source text | Check whether subject, object, status and effective scope match the source |
Three common errors
- Treating a pending review status as approved. The commission setting sheet review status is pending review, while the draft sentence states that settlement is complete. Check: input the commission setting sheet source text, compare the value of the review status field, and observe whether the result is approved or effective; when it is pending review, record the evidence gap.
- Treating an agreed agreement ratio as the basis for executed settlement. The agreement ratio and the setting sheet ratio match, but settlement execution requires the setting sheet to be effective. Check: input the agreement source text and the setting sheet source text, compare the ratio field and the effective field in both documents, and observe whether the effective field states that it is effective.
- Treating retrieval output as an authorisation record. A citation array returned by retrieval does not constitute approval or execution authorisation. Check: input the citation array returned by retrieval, compare whether it contains an approval status field, and observe whether an authorisation conclusion appears; when the field is absent, keep the matter as unverified.
How to verify the setup
- Verification one: input source text one and the draft sentence, compare the agreement signing date, the subjects and the ratio field, and observe whether the draft sentence matches the source. The expected result is that the ratio matches and the signing fact is supported, while the settlement completion conclusion is not supported.
- Verification two: input source text three, compare the value of the review status field, and observe whether it is approved. The expected result is pending review, and the settlement completion conclusion in the draft sentence remains unsupported.
- Verification three: input the citation array returned by retrieval, compare the array length with the downstream answering stage's handling design, and observe whether the downstream path still completes when the array is empty. The expected result is that the path completes at length zero, that the answering stage handles the empty citation set according to its design, and that no authorisation or settlement conclusion is produced.
A paper input of an empty array can be used to check the downstream design. This does not assert that a rule-based retrieval will return an empty array. Retrieval checks describe observed behaviour and failure branches; they do not guarantee a first-hit result, full recall, or that an unrelated query returns nothing. Optional paths shown in the overall retrieval flow are not executed on every request.