What the documents look like
Seller ratings and experience-score maintenance follows post-fulfilment consumer feedback rather than promotion pricing. The working materials are negative-review handling records, experience-score breakdowns, service diagnostic reports and rectification ledgers, drawn from platform experience scores, marketplace service indicators, after-sales review data and customer-service systems. These materials update daily or weekly, so the same draft sentence can rest on a different document version at a different time.
For copy drafting, the task is narrow: a promotional sentence placed in product or marketing copy must point back to specific wording in one of these materials. Citation tracing is the step that connects a drafted claim to its source text. The materials record warranty terms, price periods, handling status and rectification progress, and those records are the only basis for judging whether a draft sentence is supported.
Constraints at the citation tracing step
The knowledge-base search node first confirms which knowledge bases are selected and which document versions they hold. Only knowledge bases sharing the same vector model can be associated together. The node outputs a citation array, and the length of that array may be zero. A zero-length array means the chain ran through without retrieving content; it does not mean the workflow terminated with an error. Retrieval results then pass to the answering stage, where the answer and the cited source text are compared item by item. A traceable sample is the instrument for checking displayed source text; answer fluency and the model's thinking process are not substitutes for source evidence.
Similarity and rerank scores describe how closely content relates to a query. They do not perform permission or entity isolation, and retrieval itself is not authorization. In hybrid retrieval, RRF is the process that merges multiple result paths; it is not a switch that can be turned on. Reranking requires a rerank model and must be enabled in the settings. Image retrieval requires image input and a model with the corresponding capability. Version currency is checked by testing a source against current and historical documents, and whether a citation is adopted still depends on review against the effective version.
Hypothetical exercise sample: the source material reads, "From 1 March 2025, this store provides a two-year whole-machine warranty for the X model rice cooker; the warranty scope follows the terms listed on the enclosed warranty card. This model was sold at a uniform retail price of 299 yuan from 1 March 2025 to 30 June 2025." The draft sentence to be judged is, "X model rice cooker, two-year whole-machine warranty, long-term low-price supply." The objects of comparison are the warranty period, the price execution period and the supply promise; the subject is the X model rice cooker, and the effective range is the date interval above.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Comparison method and reason |
|---|---|---|
| Associated knowledge bases | Two ratings-maintenance knowledge bases selected | Check the vector model identifier first; associate only when identical, exclude otherwise |
| Search mode | Hybrid retrieval | The draft sentence contains a model name and dates, so semantic and keyword hits both need coverage |
| Citation limit | A tokens value set for this exercise | Compare against the sample source length and observe whether citations are truncated |
| Minimum relevance | A 0-1 value set for this exercise | Compare against each returned citation and observe whether low-relevance fragments are filtered |
| Result reranking | Enabled per settings | Enable when a rerank model is available, then compare rerank scores against the threshold |
| Citation handoff | Citation array passed to the answering stage | Check whether each answer sentence corresponds item by item to cited source text |
Three common errors
- Treating a retrieval score as a permission or entity filter. The input is one retrieval request; the comparison is between the score and permission identifiers; the observable result is that a high or low score has no correspondence with whether a record falls inside the target scope.
- Treating an empty citation array as a workflow error. The input is a retrieval that returns no content; the comparison is between the output array length and the node execution state; the observable result is an array length of 0 while the chain continues.
- Treating a supply promise in the draft sentence as already stated in the source. The input is the sample draft sentence and the source material; the comparison is between "long-term low-price supply" and the source price period; the observable result is that the source carries only a dated price execution statement, and the supply promise has no support.
How to verify the setup
- Warranty period check. The input is the sample draft phrase "two-year whole-machine warranty" and the warranty sentence in the source material. The comparison is between the draft wording and the period and scope terms recorded in the source. The condition is met when the draft wording shares the subject, the field and the value with the source; otherwise it is recorded as pending verification.
- Price period check. The input is the draft phrase "long-term low-price supply" and the price sentence in the source. The comparison is between "long-term" and "1 March 2025 to 30 June 2025". The condition is met when the draft wording falls inside the period recorded in the source; a period beyond that range is recorded as unsupported.
- Empty-citation handoff check. The input is a paper-specified citation array of
[]. The comparison is between that array and the downstream answering stage's handoff design. The condition is met when the downstream stage still executes as designed under an empty array. This check is a paper exercise and does not claim that an actual retrieval will return an empty array.