What the documents look like
Creator-led sales campaigns depend on external creator fulfilment and commission settlement. The working materials are creator cooperation agreements, commission setting forms, sample dispatch forms and live-stream performance reports. During an operations handover these materials commonly sit at different stages at the same time. An agreement may be signed while the cooperation has not been concluded. A commission setting form may exist only as a version awaiting review. A sample dispatch form may record dispatch alone. A live-stream performance report may cover a partial period. The retrieval and recall step has to place the material text, the handover request and the authorization state side by side for comparison. A retrieval hit is not evidence that an authorization has taken effect.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The identifiers should be checked first, non-matching bases excluded, and the actual observation recorded.
Search mode should be compared across semantic, full-text and hybrid retrieval for the test question. No single mode is assumed to suit every material. Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval fuses multiple result paths through RRF. RRF is a fusion process and is not written as a switch. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold. It does not perform permission or entity isolation. The reference limit is measured in tokens, and materials of different lengths should be used to test whether content remains complete.
Question optimization performs coreference resolution and question expansion, and it adds one model call. Image retrieval requires a corresponding image input and model capability. Reranking requires a model and must be enabled in the settings. The node outputs reference content as an array, and the length can be zero, so the empty-result path still runs through.
Hypothetical paper exercise sample: the material text reads "Party A and Party B signed the Creator Campaign Cooperation Agreement on 1 March 2025, agreeing a 15 March 2025 campaign and a 20% commission rate. Party B has not yet confirmed the schedule." The draft sentence to be assessed reads "Party B has confirmed the 15 March 2025 campaign schedule, and commission runs at 20%." The check compares the signing date, campaign date, commission rate and schedule confirmation state. The original text records only the signing and the agreed terms. It does not record schedule confirmation, so the "has confirmed" wording in the draft is not supported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more bases that use the same vector model | First check whether agreement, commission, sample and report bases share a model |
| Search mode | Compare semantic, full-text and hybrid retrieval by question type | Dates and rates suit full-text or hybrid retrieval |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant samples before adjusting the threshold |
| Reference limit | Unit is token | Test content completeness with materials of different lengths |
| Question optimization | Performs coreference resolution and question expansion | Compare the subject in the original question and the optimized question |
| Result reranking | Requires a model and must be enabled in the settings | Secondary ordering when candidates are numerous |
For this exercise, the starting value for minimum relevance is set at 0.5. The reason is that this value lies between 0 and 1 and allows the filtering difference between relevant and irrelevant samples to be observed at the same time. The comparison method is to enter the labelled relevant sample and the labelled irrelevant sample separately and observe which results are retained. This value is not an official default.
Three common errors
- Treating a retrieval hit as an effective authorization. The input is the draft sentence "Party B has confirmed the schedule." The comparison object is the schedule confirmation state in the agreement text. If the original text does not record confirmation, the draft sentence is not supported.
- Treating a pending review as a missing operation permission. The input is the pending-review state of the commission setting form. The comparison object is a completed review record. If only a pending-review marker exists, the review outcome remains to be verified.
- Associating knowledge bases that use different vector models. The input is the vector model identifiers of two bases. The comparison object is whether the identifiers match. If they do not match, the non-matching base is excluded and cross-model retrieval is not performed.
How to verify the setup
- Enter the material text and the draft sentence from the hypothetical sample. Compare "has confirmed" in the draft sentence with the schedule confirmation state in the original text. If the original text does not record confirmation, the wording is judged unsupported and the evidence gap is recorded.
- Enter an empty array [] as the reference content. Compare whether the downstream design accepts an output of length zero. If the downstream design can handle the empty array, the empty-result path runs through.
- Enter the vector model identifiers of two bases. Compare whether the identifiers match. If they do not match, exclude the non-matching base and record the actual observation without inventing a runtime result.