What the documents look like
Seller ratings management for platform stores centers on consumer feedback after fulfillment, drawing on Douyin Shop experience scores, Tmall service indicators, JD after-sales reviews and customer service systems. The day-to-day documents are negative-review handling records, experience-score detail sheets, service diagnostic reports and remediation ledgers. In store service and internal training, these materials support script drafting, responsibility attribution and remediation tracking. Citation tracing determines whether the facts behind training content can be checked line by line.
The update rhythm is daily and weekly, so version differences among documents directly affect whether a citation stays valid. A training deck that quotes last week's remediation ledger may describe a state that the current version has already changed.
Constraints at the citation tracing step
The knowledge base search 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. When nothing is retrieved, the output path still completes. The citation cap is controlled by tokens. The minimum relevance value takes effect only under semantic retrieval or result reranking. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled by configuration. Image retrieval requires image input and matching model capability. The retrieval output then passes to the answering step, where each answer sentence is checked against the cited source text. Retrieval and relevance scores carry no permission or entity-isolation duty. Knowledge bases built on different vector models cannot be associated together; check identifiers first and exclude those that do not match, then record what is actually observed.
Hypothetical drill sample: the source material reads, "From 1 March 2025 to 31 March 2025, entry 12 of a store's negative-review handling record states: the consumer reported a missing gift; customer service registered a replacement on 5 March; the replacement was completed on 8 March." A draft sentence awaits judgment: "All negative reviews at this store in March have completed remediation." The subject needed for judgment is the store, the object is the negative-review handling record, the period is March, and the field is remediation status. The source states the replacement completion for one review only. It does not state the remediation status of all reviews. The draft sentence is unsupported, and the evidence gap must remain open. It cannot be judged as completed remediation.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this drill | Check method |
|---|---|---|
| Associated knowledge bases | Select store-rating knowledge bases already confirmed to share the same vector model | Compare knowledge base identifiers and vector model identifiers; exclude mismatches |
| Search mode | Hybrid retrieval | Observe whether semantic recall and full-text recall results are merged through RRF |
| Citation cap | Set by tokens; the starting value needs a stated reason | Compare with this material's length and observe whether citations are trimmed |
| Minimum relevance | Effective only under semantic retrieval or result reranking | Compare with rerank scores and observe whether low-score content is filtered |
| Citation output handoff | Retrieval output passes to the answering step | Check each answer sentence against the cited source text; treat an empty array as zero length |
| Material currency | Test with the current version and a historical version | Compare version identifiers and confirm whether citations are reviewed against the valid version |
Three common errors
- Treating a retrieval hit as authorized execution. Retrieval returns citation content only; it does not constitute remediation of a negative review, a replacement shipment, or a permission grant. Compare the retrieval output with the authorization record. Where the authorization record is absent, record the gap.
- Extending the same-vector-model requirement to external systems. That requirement applies to knowledge base association only, not to business-registration websites or other external systems. Check only the knowledge base identifier and the vector model identifier.
- Writing an unstated remediation status as completed. The sample source states replacement completion for one review; the draft sentence claims all remediation complete. Subject and field scope differ between them. Compare the source period, object and draft sentence scope. Where scope does not match, the claim is unsupported.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare the store, period, review entry and remediation status stated in the source against the draft sentence's claim that all negative reviews completed remediation. Where the source does not state the status of all reviews, judge the draft unsupported and keep the evidence gap open.
- Input a retrieval output explicitly marked as an empty array. Compare whether the downstream answering step handles it as zero length. Observe whether the path continues to complete. Do not claim that a fictional rule search necessarily returns an empty array.
- Input two knowledge bases whose vector model identifiers differ. Compare the knowledge base identifiers and vector model identifiers. Exclude the non-matching one and record what is actually observed. Do not run a cross-model retrieval to compare missed recalls, and do not fabricate runtime results after a shared association.