What the documents look like
Creator-exclusive sales campaigns revolve around external creator fulfilment and commission settlement. Brand commercial teams, MCN agencies and creator managers jointly handle creator cooperation agreements, commission setting sheets, sample dispatch notes and livestream performance reports. The comparison task places agreement terms and transaction records on the same field, for example commission rate, settlement period and slot arrangement. Materials update per session and per week; the retrieval step organises the existing document scope and does not derive batch synchronisation or transaction-triggered mechanisms.
A paper exercise sample is used throughout this page. Material text one: "Party A commissions Party B to run an exclusive livestream on 18 June 2025, with a commission rate of twenty percent of transaction value, and samples to arrive before 10 June." Material text two: "The commission setting sheet states the commission rate for this session is fifteen percent, and the settlement period is thirty days after the livestream ends." The draft sentence to be judged: "This session settles commission at twenty percent, with a settlement period of fifteen days after the livestream." This sample is for paper comparison only and points to no real entity.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together; the operator first checks identifiers, excludes non-matching bases, and records actual observations, without running cross-model retrieval to compare missed recalls. The search mode is compared across semantic search, full-text search and hybrid search according to the test question; no single mode is presumed suitable for all materials. The reference cap is measured in tokens, and materials of different lengths are used to test whether content is complete. The minimum relevance score takes effect only under semantic search or when result reranking is used; relevant and irrelevant samples are labelled before adjusting the threshold. Question optimisation performs coreference resolution and question expansion, and the original question is compared with the optimised question to check whether the subject was altered. Reference content is output as an array and its length can be zero; when nothing is retrieved, the output path still completes. Retrieval and similarity scores do not carry permission or entity isolation.
For the paper sample, the agreement states twenty percent and the commission setting sheet states fifteen percent for the same field, which is a same-field mutually exclusive value pair and is judged as a contradiction, pending verification. The settlement period of thirty days in the setting sheet and fifteen days in the draft sentence are likewise a same-field mutually exclusive pair. Fields not stated in the materials are recorded as unsupported, and no authorisation or effectiveness clause is supplied to reach a determinate answer.
Configuration and implementation checks
| Configuration item | Setting applicable to this page | Correspondence with the sample |
|---|---|---|
| Associated knowledge bases | Associate one or more knowledge bases using the same vector model | Agreement and commission setting sheet must sit within the same vector model scope |
| Search mode | Compare semantic, full-text and hybrid search per test question | Commission rate and settlement period fields are checked with full-text or hybrid search |
| Reference cap | Unit is token; test with materials of different lengths | Covers both material texts and checks whether content is complete |
| Minimum relevance score | Takes effect only under semantic search or when result reranking is used | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimisation | Used for coreference resolution and question expansion | Compare original and optimised question, check the subject was not altered |
| Result reranking | Reranks candidates when a model is available and enabled per settings | Used to compare relevance scores when candidates are numerous |
The reference cap starting value for this paper exercise is set to a level that covers both material texts in full; the reason is that the two texts together form the comparison basis, and the value is compared against the combined token length of the two texts to observe whether truncation occurs. This is a hypothetical starting value for the exercise and is not an official default. The minimum relevance score starting value is not fixed here; relevant and irrelevant samples are labelled first, and the threshold is compared against those labels.
Three common errors
- Treating the "twenty percent" in the draft sentence as a confirmed settlement basis. The agreement text and the commission setting sheet give different values for the same field, which is a same-field mutually exclusive pair, and is judged as a contradiction pending verification; which document prevails is not inferred.
- Merging "settlement period is thirty days after the livestream ends" with the draft sentence's "fifteen days" into one fact. Both belong to the settlement period field with mutually exclusive values, which is a contradiction; parts not stated are recorded as unsupported.
- Treating a retrieval hit as completed authorisation or approval. Retrieval results only provide reference content; business verification, approval, database writes and external synchronisation require their respective downstream implementations, and a pending review does not demonstrate the absence of operating permission.
How to verify the setup
- Input the two material texts and the draft sentence from the paper sample, and compare the commission rate field: the agreement gives twenty percent, the commission setting sheet gives fifteen percent, and the draft sentence gives twenty percent. When same-field mutually exclusive values appear, judge as a contradiction and retain the evidence gap. The expected observation is that the comparison surfaces both values rather than selecting one.
- Input the same material set and compare the settlement period field: the commission setting sheet gives thirty days and the draft sentence gives fifteen days. When the two are mutually exclusive, judge as a contradiction; other clauses not stated in the materials are recorded as unsupported, and no authorisation or effectiveness clause is fabricated. The expected observation is that the unsupported clauses remain marked as gaps.
- Input a follow-up question with unclear reference, compare the original question with the retrieval terms after question optimisation, and check whether the subject was altered. Then use the explicit paper input
[]to check whether the downstream path still completes on an empty result, without claiming that a fictional rule retrieval necessarily returns an empty array. The expected observation is that the output path completes and the empty reference is handled as designed.