What the documents look like
Creator-led sales campaigns depend on external creator fulfillment and commission settlement, which separates this work from brand-operated live rooms that rely on internal staffing. The materials handled each campaign and each week include creator cooperation agreements, commission setting sheets, sample shipment sheets, and live campaign reports. Each material carries a distinct fact. An agreement records signing and effectiveness status. A commission setting sheet records the rate and the settlement basis. A sample shipment sheet records dispatch and receipt. A live campaign report records the transaction result of a session. The retrieval and recall step organizes these materials by knowledge base so that order fulfillment and after-sales answers can draw the corresponding source text. Retrieval results reflect only the index content that existed at the time of ingestion. A retrieved agreement passage does not show that cooperation was reached, that a contract took effect, or that any authorization action occurred.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The identifier of each base should be checked first, and any base whose vector model does not match should be excluded and the actual observation recorded; a cross-model retrieval should not be run to compare missed recall. Search modes include semantic retrieval, full-text retrieval, and hybrid retrieval, and these should be compared against the same test question rather than assuming one mode suits all materials. The citation limit is measured in tokens. The minimum relevance applies only during semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a toggle. Image retrieval requires a corresponding image input and model capability, and reranking requires an available model enabled under its settings. Retrieval and recall do not perform authorization judgment.
Hypothetical paper exercise sample: in one creator campaign, the source text provided for checking reads "This agreement takes effect on the date of signature by both parties, and the signing date is March 10, 2025," "The commission setting sheet states that the commission rate for this session is fifteen percent, and the settlement basis is settlement after confirmed receipt," and "The sample shipment sheet states that the sample was dispatched on March 5, 2025." A draft sentence to be assessed reads "The creator cooperation took effect on March 10, 2025, and commission is settled at fifteen percent." Assessment requires the same subject, the same period, and the same field. The agreement effectiveness date matches the draft sentence and can be supported. The commission rate matches the setting sheet and can be supported. Whether settlement has been completed is not stated in the materials and remains unsupported. Dispatch and receipt are two separate facts, and a dispatch date does not show receipt.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Agreement base, commission setting sheet base, sample shipment sheet base, live campaign report base | Check the vector model identifier of each base first, exclude mismatched ones, then decide the shared association scope |
| Search mode | Hybrid retrieval | Compare recall content for the same test question against semantic retrieval and full-text retrieval separately |
| Citation limit | Set to 2000 tokens as the hypothetical trial starting point | Test with source materials of different lengths to see whether cited content is complete, then compare against higher or lower values |
| Minimum relevance | Set to 0.5 as the hypothetical trial starting point | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled under its settings | Confirm the rerank model is available and compare candidate ordering before and after reranking |
These values are hypothetical trial starting points used to explain the comparison method and are not official defaults. The citation limit unit is the token. Minimum relevance is a filtering threshold, not a final ordering rule, and it does not perform permission or entity isolation.
Three common errors
- Treating a retrieval hit as authorized execution. The input is the draft sentence "The creator cooperation took effect on March 10, 2025," and the agreement text is retrieved and then treated as showing cooperation was reached. The comparison is between the signing date stated in the agreement and the date in the draft sentence, and the observable result is that a matching date can be supported. Cooperation being reached and a contract taking effect are separate facts, and a gap is recorded when the materials do not state them.
- Treating a pending review as a lack of operating permission. The input is a handover request stating "the commission setting sheet is pending review," and this is used to conclude that the session has no settlement permission. The comparison is between the review status stated in the material and the permission field, and the observable result is that an unstated permission field is unsupported and cannot be called a denial.
- Associating knowledge bases built on different vector models. The input is several knowledge base identifiers associated without checking the vector model. The comparison is between the vector model corresponding to each base identifier, and the observable result is that mismatched identifiers should be excluded and recorded, without inferring runtime behavior.
How to verify the setup
- Input the agreement source text and the draft sentence from the hypothetical sample, compare whether the signing date stated in the agreement and the date in the draft sentence share the same subject, the same period, and the same field, and observe that a matching date supports the sentence while an unstated date is recorded as unsupported.
- Input the commission setting sheet source text and the draft sentence from the hypothetical sample, compare whether the commission rate and the settlement basis belong to the same field, and observe that a matching rate can be supported while whether settlement is complete is not stated in the materials and is recorded as pending verification.
- Input a clearly written paper empty result [], check how the downstream design handles an empty citation, and observe that the output path still runs when the citation length is 0; no claim is made that a fictional rule retrieval necessarily returns [].