What the documents look like
Short-video affiliate distribution relies on long-term content exposure and product-link conversion. It does not depend on live explanation. The main data sources are the selected affiliate program, short-video playback data, product transaction data and platform rules, updated daily and weekly. Everyday documents include short-video scripts, product lists for attached links, commission plans and violation feedback.
For order fulfillment and after-sales question answering, retrieval needs concentrate on three content types: which products a given video carries, the commission plan terms tied to a given order, and the processing status of a given violation feedback record. These documents differ in subject, object and effective scope. Treating a product mention in a script as the attached-link product list, or treating a commission plan draft as an effective clause, will produce wrong answers in after-sales replies.
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. During operation, first check the model identifier, exclude non-matching bases and record the actual observation.
Retrieval output is an array of quoted content, and its length can be zero. When nothing is found, the output path still completes. Semantic retrieval, full-text retrieval and hybrid retrieval are compared against test questions; no single mode is assumed to suit all material. The citation limit is measured in tokens and controls the maximum amount of content quoted per search. Minimum relevance is a 0-1 filter threshold that takes effect only in semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion; the original question and the optimized question are compared. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled by configuration. Optional paths shown in the overall flow diagram do not mean every request executes all of them. Retrieval and recall do not perform authorization judgment; base scope and similarity scores do not constitute permission or entity isolation.
Hypothetical paper exercise sample: the source text reads, "Video A was published on June 3, carrying SKU-1001 and SKU-1002. Commission plan P-07 has applied to SKU-1001 since June 1 at a rate of 12%. Violation feedback F-22 concerns the title of Video A and its status is pending review." The draft sentence to be assessed is: "Commission for SKU-1001 on Video A has been settled at 12%, and violation feedback F-22 has been fully processed." Comparing this sentence with the source: the settlement action is not stated in the source, and the status of F-22 is pending review. Pending review does not constitute fully processed.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Attached-link product list base and commission plan base; verify that vector model identifiers match | Both bases need the same vector model to be associated together; check identifiers first, then exclude non-matching bases |
| Search mode | Start with hybrid retrieval | Commission plans contain identifiers such as P-07, where full-text recall has an advantage; compare results item by item against pure semantic retrieval |
| Citation limit | Set to 2000 tokens | Starting point for this exercise because commission clause passages are long; compare content completeness against a 1000-token result |
| Minimum relevance | Set to 0.4 | Starting point for this exercise; label relevant and irrelevant samples before adjusting; takes effect only in semantic retrieval or reranking |
| Question optimization | Enable in multi-turn follow-up scenarios | Used for coreference resolution; compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enable when candidates are numerous | Requires a reranking model and must be enabled by configuration; compare filtering by rerank score against filtering by vector score |
Three common errors
- Reading a pending-review status as fully processed. The input is "violation feedback F-22 has status pending review" in the hypothetical sample. The comparison object is the draft sentence "F-22 has been fully processed." The observable result is that the status field values differ, and the draft is judged unsupported.
- Extending a commission plan's effective scope to products not listed. The input is "P-07 has applied to SKU-1001 since June 1." The comparison object is the draft sentence's settlement assertion for SKU-1001. The observable result is that the source does not state the settlement action, and the draft is judged to have an evidence gap.
- Treating a retrieval score as a permission filter. The input is a citation array returned by one semantic retrieval. The comparison object is the authorization scope list. The observable result is that score level has no correspondence with authorization status, and retrieval is judged not to carry authorization.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare whether the settlement assertion in the draft sentence matches the source in subject, period and field. The observable result is that when the source does not state the settlement action, the draft is judged unsupported, not judged as a negation or conflict.
- Input one retrieval request targeting P-07. Compare whether the returned citations contain the original P-07 clause text. The observable result is that the condition is met when the citation array contains that clause. When the result is empty, use an explicit paper input to check the downstream design, and do not claim that a fictional-rule retrieval necessarily returns an empty array.
- Input one request to associate bases built on different vector models. Compare the vector model identifiers of the two bases. The observable result is that when identifiers do not match, that base is excluded and recorded, and cross-model retrieval is not executed to compare missed recalls.