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 working materials are short-video scripts, product lists for attached links, commission plans, and violation feedback. The brief gives a daily or weekly update frequency; that frequency does not change the citation-tracing function described here, so it is noted and set aside.
For order fulfillment and after-sales question answering, citation tracing answers a narrow question: which script, which product list, or which commission plan supports a given fulfillment statement or after-sales reply, and whether that supporting material is still valid for the order period. When material versions and order periods do not line up, a reply can cite an expired position.
The category differs from live-room selling in that the content keeps working after publication. A reply drafted today may rest on a script written weeks earlier. The citation record is the only way to connect the two.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. The model identifier is checked first, and any base that does not match is excluded. Cross-model retrieval is not run to compare missed recalls.
The search node outputs cited content as an array. The length may be zero. An empty array does not mean the workflow has errored and stopped; downstream handling can be checked with a paper input of [].
Retrieval output is passed to the answering step, where each answer sentence is checked against the cited source text. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or result reranking. Relevance is a filter threshold; it does not carry permission or entity isolation. Retrieval itself is not authorization. RRF is a fusion process for mixed results, not a switch. Reranking, image retrieval, and question optimization each require the corresponding model capability and must be enabled by configuration. The optional paths in the overall document flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: the source text reads "Product list for attached links, Version A, applicable period 1 March 2025 to 31 March 2025; commission plan Version B valid over the same period; violation feedback Record C, dated 12 March, title wording pending correction." One draft sentence to be judged reads "This order's after-sales handling follows Version A list and Version B plan; Record C has been corrected." Record C's status is pending correction. The draft sentence states correction is complete, which does not match the source status, so that sentence is unsupported. Whether Version A list and Version B plan cover the order date must be checked against the order date field. Where that field is missing, the gap is recorded.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Check method and reason |
|---|---|---|
| Associated knowledge bases | Product list base, commission plan base | Check that both bases share the same vector model identifier; exclude any base that does not match, and do not merge retrieval across models |
| Search mode | Hybrid retrieval | List numbers and plan names contain proper terms; semantic and full-text paths are combined through RRF |
| Citation limit | Trial starting point of 2000 tokens | This is a trial starting point only, compared against the sample source length and adjusted; it is not an official default |
| Minimum relevance | Trial starting point of 0.3 | This is a trial starting point only, compared against the score distribution after reranking is enabled |
| Result reranking | Enabled | A rerank model must already be configured; when it is not configured, reranking is not enabled |
| Cited content handling | Output array passed to the answering step | Use the sample source text to check each answer sentence against the cited content |
Three common errors
- Treating a retrieval hit as authorization. Input: the after-sales request and the retrieval result. Compare: the retrieval output against the authorization material. Observe: when retrieval returns a result but no corresponding authorization record exists, the request is still handled as unauthorized.
- Writing pending correction as completed correction. Input: the draft sentence and Record C. Compare: the status value of the same field. Observe: pending correction and completed correction are mutually exclusive values; when both appear for one field, the sentence is inconsistent with the source.
- Forcing association of knowledge bases with different vector models. Input: the model identifiers of the two bases. Compare: whether the identifiers are the same. Observe: when the identifiers differ, the base is excluded and cross-model retrieval is not run.
How to verify the setup
- Input the sample source text and the draft sentence "Record C has been corrected." Compare the status field of Record C. Observing that the status is pending correction satisfies the condition that the sentence is unsupported.
- Input a paper citation array of []. Compare the downstream answering step design. Observing that the downstream step continues without error on an empty citation satisfies the empty-result handling condition.
- Input the model identifiers of the product list base and the commission plan base. Compare whether the two identifiers are the same. Observing that they match satisfies the condition for joint association; when they differ, record the difference and exclude the base. All of the above are paper checks. No retrieval or business action has actually been executed.