What the documents look like
Short-video affiliate sales relies on long-term content exposure and product-link conversion rather than live explanation. The working materials for this category include short-video scripts, attached-product lists, commission plans, and violation feedback, with updates occurring daily or weekly. For contract and transaction material comparison, each script, list, and commission plan must be checked field by field against the cited source text: the parties, the product, the ratio, and the validity period in a draft sentence need to be traceable to the material that is said to support them. The retrieval step only locates candidate citations. Authorization status, signing facts, and effective scope remain facts that the underlying materials themselves must state.
The retrieval node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array means the output path still completes; it does not mean the workflow has errored or terminated. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a rerank model and must be enabled according to settings. The minimum relevance value is a filtering threshold and does not perform permission or entity isolation. The citation limit is measured in tokens. Retrieved results must be passed to the answering step before the draft sentence can be checked against the source text line by line.
Hypothetical paper exercise sample: the source material reads, "Party A authorizes Party B to promote attached-product A from March 1, 2025 to June 30, 2025, at a commission rate of fifteen percent; this authorization takes effect on the date both parties sign, and the signature date field is blank." The draft sentence to be judged is, "Party B obtained promotion authorization for product A on March 1, 2025, at fifteen percent commission." The comparison targets the authorization period, the commission rate, and the effective condition; the parties are Party A and Party B, the object is product A, the status is a blank signature date, and the effective range ends June 30, 2025. The draft sentence treats the start of the authorization period as an effective fact, while the material does not state that signing was completed, so the sentence is unsupported.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that use the same vector model. When several bases are selected, their vector-model identifiers must be checked first; bases whose identifiers differ cannot be associated together. This check is a comparison of model fields, not a retrieval run across mismatched models.
The citation output is an array and may contain zero entries. A zero-length array is a valid output of the search node and continues through the downstream path. Treating it as a failure would misread the documented behavior. Downstream design should be examined to see whether an empty array is handled as a normal input.
Hybrid retrieval combines semantic recall and full-text recall, and RRF is the merge process for those result paths. RRF is not a separate switch to be turned on. Reranking requires a rerank model and must be enabled according to the configured settings; without the model, reranking does not occur. The minimum relevance value applies only under semantic retrieval or when result reranking is used, and it filters low-relevance results. It does not exclude unrelated businesses or unauthorized records, because relevance scoring is not a permission filter.
The citation limit is expressed in tokens per search. This is different from the thinking budget used in AI configuration, which controls reasoning effort and is not a citation token limit. The two values must not be substituted for each other.
The overall retrieval flow lists several optional paths. Image retrieval requires image input and a model capable of handling it. Reranking requires a rerank model and must be enabled. Not every request executes every listed path. When a path is not configured or its input is absent, that path does not run.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select only bases containing the authorization letter and commission plan that share one vector model | Check vector-model identifiers first; exclude bases with different identifiers; compare the model field across candidate bases |
| Search mode | Hybrid retrieval | The authorization letter contains exact fields such as dates and ratios; compare semantic recall and full-text recall results |
| Citation limit | A starting value set to cover the full text of both materials in tokens | This is a starting value for this exercise; compare it with the actual material length and adjust |
| Minimum relevance | A lower starting value | This is a starting value for this exercise; compare it with whether returned citations cover the authorization clauses |
| Result reranking | Enabled and rerank scores checked | A rerank model must be available; compare candidate order before and after reranking |
| Citation output handoff | Check whether the citation array is passed to the answering step | Compare each answer sentence with the cited source text line by line |
Three common errors
- Treating a retrieval hit as an effective authorization. Input is the authorization letter text above; compare the draft sentence with the blank signature date field; observe whether the start of the authorization period is written as an effective fact.
- Treating minimum relevance as a permission filter. Input is an unrelated product record inside the same knowledge base; compare the returned citations with the authorized object; observe whether the threshold is assumed to have excluded unauthorized content.
- Treating an empty citation array as a workflow failure. Input is a paper citation array of []; compare the downstream node design; observe whether the empty array is recorded as an error or termination.
How to verify the setup
- Input the authorization letter text above and the draft sentence; compare the authorization period, commission rate, and effective condition fields; the condition is met when the draft sentence has no material support on the effective condition.
- Input an unrelated product record inside the same knowledge base; compare the returned citations with the authorized object; the check condition is met when the returned citations do not contain that product and the actual observation is recorded.
- Input a paper citation array of []; compare the downstream handoff design; the check condition is met when the downstream path continues with the empty array and does not record an error. All three items describe expectations only and were not actually executed.