What the documents look like
Seller ratings and experience-score maintenance draws on consumer feedback that arrives after fulfillment. The public brief lists Douyin Shop experience scores, Tmall service indicators, JD after-sales reviews and customer-service systems as the main sources. The documents handled day to day are negative-review handling records, experience-score detail sheets, service diagnostic reports and rectification ledgers, with daily and weekly update rhythms. For this page the task is product-information and selection questions: retrieving content about product attributes, service wording and indicator definitions so that operations staff, customer-service supervisors and quality inspectors can check it. Retrieval returns stored text fragments only. It does not decide authorization, grant approval or write to a database.
Two characteristics shape this task. First, the material mixes structured indicator values with free-text handling notes, so a question may contain both a product identifier and a natural-language description. Second, the records carry states such as pending supplementary evidence, which readers can mistake for a settled conclusion. The update rhythm does not change the retrieval step itself, so it is not developed further here.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval; hybrid retrieval merges multiple result paths through RRF. Result reranking uses a ReRank model to sort candidate text a second time. The citation limit controls, in tokens, how much content one search may cite. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion before retrieval and adds one model call. Image retrieval requires image input and the corresponding model capability, and reranking requires a model and must be enabled by configuration. Optional paths shown in the overall flow diagram are not all executed on every request. Retrieval and similarity scores are not permission filters and do not exclude unrelated businesses or unauthorized records.
Hypothetical paper exercise. A shop operator submits a handover request on 10 June 2025: "Using the May 2025 negative-review handling records, confirm the after-sales compensation wording for Product A and produce a selection-question draft." The source material supplied for checking reads: "May 2025 negative-review handling record, item 12: Product A received a negative review because of a logistics delay. Customer service contacted the consumer on 18 May to request supplementary evidence. The consumer did not reply. This record's status is pending supplementary evidence. The May 2025 experience-score detail sheet shows Product A's service indicator at 4.6, against a shop average of 4.7 for the same period." The draft sentence to be judged is: "The after-sales compensation wording for Product A is confirmed and may be given externally." The draft sentence matches the source material in subject, object and period. The source material does not state that compensation wording is confirmed, so the draft sentence is unsupported.
Configuration and implementation checks
The table below lists documented settings or implementation checks relevant to this feature and task, tied to the exercise above. The values are hypothetical starting points for this exercise only, used for comparison, and are not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the ratings-maintenance knowledge base | That base uses one vector model, meeting the node's same-vector-model requirement; compare vector-model identifiers before adding another base |
| Search mode | Hybrid retrieval | The exercise question contains a product identifier and natural-language description; hybrid retrieval uses semantic and full-text recall together; compare recalled content against semantic-only and full-text-only runs |
| Citation limit | Hypothetical starting point of 2000 tokens | A single negative-review record is long, so the value is chosen to hold a complete record; compare against 1000 tokens to see whether cited content is truncated |
| Minimum relevance | Hypothetical starting point of 0.5 | Takes effect only under semantic retrieval or reranking; compare against 0.3 and 0.7 by the proportion of relevant and irrelevant fragments returned |
| Question optimization | Enabled | The exercise request contains references and omissions; compare the original question and the optimized question to check whether the subject was altered |
| Result reranking | ReRank enabled | Sorts text a second time when candidates are numerous; compare ordering of the same question before and after reranking |
Three common errors
- Treating a retrieval score as a permission filter. Input: the exercise request and the ratings-maintenance knowledge base. Compare whether returned content includes other shops or unauthorized records. The observable result is that retrieval does not perform authorization checks, so access scope must be verified separately.
- Writing pending supplementary evidence as confirmed compensation wording. Input: item 12 of the source material and the draft sentence. Compare the value each assigns to the same field. The observable result is that the source material does not state compensation wording, so the draft sentence is unsupported.
- Associating knowledge bases built on different vector models. Input: two knowledge bases with different vector-model identifiers. Compare the node's association condition. The observable result is that the node requires the same vector model, so identifiers must be checked first and non-matching bases excluded.
How to verify the setup
- Input the exercise request and the source material. Compare whether the returned cited content includes item 12 and the experience-score detail sheet. The observable result is that cited content covers the fields for that period and Product A, and does not contain a statement that compensation wording is confirmed.
- Input the draft sentence and the source material. Compare subject, object, period and field values. The observable result is that the draft sentence is unsupported, and the evidence gap must be recorded.
- Input a natural-language question without a product identifier. Compare content returned by semantic retrieval and by hybrid retrieval. The observable result is that the two modes return differing fragments; a first-position hit or complete recall is not guaranteed.
An empty result can be checked with a clearly labelled paper input of an empty array against the downstream design; no fictional rule is claimed to return an empty array. When vector-model identifiers differ, exclude the non-matching base and record the observation rather than running a cross-model search to compare missed recall.