What the documents look like
Short-video affiliate sales rely on long-term content exposure and product-link conversion rather than live explanation. In store service and internal training, four document types form the main checking material: short-video scripts, affiliate product lists, commission plans, and violation feedback. These documents update daily or weekly. Citation tracing at this step addresses one core question: can the rules, commission ratios, and product statuses cited in training material or service replies be matched to verifiable original documents and valid versions. Because content remains available for long periods, a citation written into training material is reused over an extended time, and the effect of a missing source label or an expired version accumulates.
The knowledge base search node selects one or more knowledge bases that use the same vector model and outputs an array of citations whose length may be zero. When the array is empty, the output link still runs. Search results pass to the answer step, where the answer and the cited original text are checked item by item. Minimum relevance is a filtering threshold that takes effect only in semantic search or when result reranking is used; it does not perform permission or entity isolation. Hybrid search merges multiple recall paths through RRF, and Rerank reorders candidate text. These are processing steps inside the retrieval chain. Document timeliness checks use the current version and a historical version to test the source separately, and whether a citation is adopted still depends on review of the valid version.
Constraints at the citation tracing step
A hypothetical drill sample is given here. A draft internal training document for one store contains the following original text: "The March 2025 version of the Affiliate Product List marks Product A as on sale, commission plan number C-0315, applicable period 1 March 2025 to 31 March 2025." "The April 2025 version of the Affiliate Product List marks Product A as delisted." The draft sentence to be judged reads: "Training material states 'Product A is currently on sale, commission executed under C-0315'." The check compares the on-sale status cited in the draft sentence with the field values in the two list versions. The on-sale status comes from the March version, while the April version already marks the product as delisted. The two versions are mutually exclusive values for the same product and the same field. The draft sentence does not label the version or the applicable period, so its source labeling is incomplete.
The same product, the same period, and the same field are required for a judgment. A clear mutual exclusion between values in the same field is a contradiction. A statement that is not recorded, or that has only evidence from a related stage, is unsupported or pending verification. Pending review does not demonstrate a lack of operating permission. Signing, reaching cooperation, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Authorization or effectiveness clauses are not to be supplied to obtain a definite answer, and evidence gaps are retained explicitly.
Material needed for a check must sit inside a target retrieval scope to which access is held, or inside a clearly identified manual input. When it is missing, the gap is recorded. A retrieval hit is not by itself a verified source. Minimum relevance is not a permission filter. An empty citation array is not a workflow failure.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Store training material library, single vector model | Check the selected knowledge base identifier and vector model identifier; only bases with the same vector model can be associated together |
| Search mode | Hybrid search | Product numbers and commission plan numbers are exact queries; hybrid search uses semantic recall and full-text recall together |
| Result reranking | Enabled | When candidate results are numerous, text is reordered a second time; citation order is checked against the rerank score |
| Citation limit | Hypothetical trial starting point: measured in tokens, specific value pending | Documents use tokens as the unit of the citation limit; the value is set against the length of the training material |
| Minimum relevance | Hypothetical trial starting point: a value in the 0-1 range, pending | This value takes effect only in semantic search or when result reranking is used; it is adjusted against actual recall behaviour |
| Citation output handling | Retrieval output connected to the answer step | The answer and the cited original text are checked item by item; answer fluency does not replace source evidence |
These entries are documented settings or implementation checks tied to the documents on this page. They are not default values, limits, or global switches. The citation limit is a token scope, not a thinking budget. RRF is a merge process for hybrid results, not a switch to be turned on. Optional paths in the overall retrieval flow do not run on every request: image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled in the settings.
Three common errors
- Treating a retrieval hit as a verified source. Input: the draft training sentence about Product A. Comparison: the cited original text returned by retrieval against the wording of the draft sentence. If the cited original text does not label the version and applicable period, the on-sale statement in the draft sentence lacks source support and is recorded as pending verification.
- Treating minimum relevance as a permission filter. Input: a commission plan query spanning stores. Comparison: whether the returned citation falls within the authorized scope of that store. Minimum relevance filters only low-relevance results; the authorized scope is checked separately.
- Treating an empty citation array as a workflow failure. Input: a retrieval request with no matching content. Comparison: whether the node output is an array of length zero. When the array is empty, the output link still runs, and the downstream answer step's handling of empty citations is checked.
How to verify the setup
- Input the March and April list original text from the hypothetical sample. Compare the on-sale status cited in the draft sentence with the field values in the two list versions. If the two versions are mutually exclusive values for the same product and the same field, and the draft sentence labels neither version, the source labeling is judged incomplete.
- Input commission plan number C-0315 and its applicable period. Compare whether the cited original text returned by retrieval contains that number and period. If the cited original text contains them and the period matches the draft sentence, that part of the source is traceable. If the cited original text is missing, record the evidence gap.
- Input an explicit paper empty result array []. Compare the downstream answer step's design for handling empty citations. If the downstream design prompts that no source was retrieved, the empty-result handling meets the expectation. If the downstream directly generates an answer without a source, record it as pending correction.
These checks state expected observations and failure branches. They do not assert that a first hit is always returned, that all relevant items are recalled, or that an unrelated query returns an empty result. The empty-result case uses the explicit paper input [] to check downstream design; it does not claim that a fictional rule retrieval returns []. Knowledge bases using different vector models cannot be associated together. Identifiers are checked first and non-matching bases are excluded; cross-model retrieval is not run to compare missed recall.