Creator sales campaigns: citation tracing for document comparison

Review citation tracing for document comparison in Creator sales campaigns. Check sources, input boundaries, error handling and repeatable test evidence.

What the documents look like

Creator sales campaigns are built around external creator fulfillment and commission settlement. Brand commercial staff, MCN agencies, and creator managers work from the same set of materials: creator cooperation agreements, commission setting sheets, sample shipment forms, and livestream performance reports. These documents come from three different source types—contract, transaction, and fulfillment—and their field definitions frequently diverge. An agreement states a commission rate, a setting sheet records the value actually placed into effect, and a performance report reflects the resulting transactions. A comparison task requires returning to each source line by line to confirm whether figures, parties, and periods correspond.

The materials update per session and per week, and versions shift as negotiation and settlement proceed. A citation must be able to point back to a specific version. Because the agreement, the setting sheet, and the report may each carry a different version date or scope, tracing a citation means identifying which version of which document supports a given statement, and recording when no version does.

Constraints at the citation tracing step

The knowledge base search node selects one or more knowledge bases that share the same vector model and outputs a citation array whose length may be zero. A zero-length array still passes through the output path; it does not by itself terminate the workflow. Retrieved citations are handed to the answering step and must be checked line by line against the source text. Retrieval is not authorization. A similarity score is not a party or permission filter, and it does not exclude unrelated businesses or unauthorized records.

Minimum relevance takes effect only under semantic retrieval or result reranking. The citation upper limit is measured in tokens. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a rerank model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. These optional paths do not all execute on every request. Knowledge bases using different vector models cannot be associated together; the model identifier must be checked and non-matching bases excluded.

Hypothetical paper test case: A creator cooperation agreement states, "Party A commissions Party B to conduct a single livestream session on June 1, 2025, with a commission rate of 15% of transaction value, settled according to shop transaction details." A commission setting sheet states, "Commission rate for this session: 15%, effective date June 1, 2025." A draft sentence to be assessed reads, "Commission for this session is settled at 20%." The comparison inputs the agreement text, the setting sheet text, and the draft sentence, then compares the same field—commission rate. Both source texts show 15%; the draft shows 20%. This is a mutually exclusive value in the same field, and the draft sentence is not supported by the source text. If the draft sentence instead reads, "Commission settlement follows shop transaction details," it matches the agreement text and is supported. If the materials do not state a settlement cycle, that field remains unsupported and the gap is recorded.

Configuration and implementation checks

Configuration or check itemHypothetical starting point for this pageReason and comparison method
Associated knowledge basesSelect bases holding agreements, setting sheets, and reportsFirst check the vector model identifier; only matching models may be associated together, and non-matching ones are excluded
Search modeHybrid retrievalFull-text recall for identifiers and proper nouns, semantic recall for natural-language phrasing, merged through RRF
Minimum relevanceHypothetical starting value 0.3This is a trial starting point only; compare against 0.5 to observe recall range changes, not an official default
Citation upper limitHypothetical starting value 2000 tokensThis is a trial starting point only; compare against 1000 to observe truncation, not an official default
Citation output handlingCitation array passed to the answering stepCheck each answer statement against cited source text; a zero-length array is handled as length 0
Material timeliness checkCurrent version and historical version held togetherUse both versions to test source pointing; adoption still depends on valid-version review

Three common errors

  1. Treating a retrieval hit as authorization. Input a query, compare the returned citation against the agreement source text. If a citation exists but the agreement does not state signing or effectiveness, that status remains unsupported and the cooperation cannot be judged as concluded. Signing, cooperation conclusion, and contract effectiveness are three distinct facts.
  2. Treating a similarity score as a party filter. Input a creator name query, compare the returned content against the agreement party field. A high score with a mismatched party is unsupported and cannot serve as a settlement basis.
  3. Forcing knowledge bases with different vector models into joint association. Input two base identifiers, compare the vector model field. If they differ, exclude them. Do not execute cross-model retrieval to compare missed recalls.

How to verify the setup

  1. Input the agreement text, the setting sheet text, and the draft sentence "Commission for this session is settled at 20%" from the hypothetical sample. Compare the commission rate field. Both source texts show 15%; the draft shows 20%. The result is a mutually exclusive value in the same field, and the draft sentence is unsupported.
  2. Input the agreement text and the draft sentence "Commission settlement follows shop transaction details." Compare the settlement basis field. The two match, so the sentence is supported. If the materials do not state a settlement cycle, that field is recorded as unsupported.
  3. Input an explicit paper citation array [] and check whether the downstream handling design passes through the output path at length 0. Do not claim that a given retrieval rule necessarily returns [].

This page addresses only creator sales campaigns, contract and transaction material comparison, and citation tracing. Operational handover must not treat retrieval or a form as an authorization action. Whether materials, operations, or requests are in an authorized state at handover remains a separate determination that the cited documents must support.