What the documents look like
Platform store rating and experience-score maintenance draws on data from Douyin Store experience scores, Tmall service indicators, JD after-sales reviews, and customer-service systems. The order-fulfillment and after-sales question-answering stage handles consumer feedback that follows after-sales fulfillment, and it does not directly determine promotional pricing. The working documents are negative-review handling records, experience-score detail sheets, service diagnostic reports, and rectification ledgers, with update frequency at daily or weekly intervals. During retrieval these materials enter the knowledge base as text fragments, and the returned references help a customer-service supervisor or quality-control specialist check facts. Retrieval itself does not change ratings or experience scores; it returns reference content only.
The retrieval step is where after-sales question-answering either finds the right fragment or does not. A negative-review record may describe a refund date and a buyer comment, while a rectification ledger may record whether an order was placed under rectification. These are separate fields in separate documents. Retrieval can surface both, and it cannot merge them into a single authorization or status.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges the two paths through RRF. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold, not the final sorting rule. The reference cap controls, in tokens, the maximum content cited per search. When question optimization is enabled by configuration, it performs coreference resolution and question expansion and makes an additional model call. The same-vector-model requirement applies to knowledge base association only and does not extend to other external systems. Retrieval results do not constitute authorization and do not isolate entities or permissions. An empty reference output may have a length of zero, and the output path still proceeds.
Several documented behaviors matter for after-sales work. Image retrieval is not a separate capability outside the knowledge base search path; it adds an image-input processing branch, and whether an image can be retrieved depends on whether a corresponding index was built at ingestion time, not only on whether an image was uploaded at query time. The overall retrieval flow combines text, image, semantic recall, full-text recall, question optimization, and reranking, then fuses multiple result paths. The optional paths in that flow do not all execute on every request: image retrieval requires image input and model capability, and reranking requires an available model enabled by configuration. RRF is the fusion process applied after hybrid retrieval, not an independent switch.
Configuration and implementation checks
The following configuration is set around the hypothetical drill sample on this page. The numeric values are starting points for this hypothetical trial, not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Associate only the same-vector-model base holding negative-review handling records and the rectification ledger | First verify that the base identifier and vector model identifier match, and exclude bases with different models; compare whether the retrieval scope covers the sample materials before and after association |
| Search mode | Hybrid retrieval | The sample contains natural-language phrasing and an order number; compare semantic retrieval and full-text retrieval recall separately, then observe the RRF-fused result |
| Reference cap | Hypothetical starting point of 2000 tokens | Used to compare the completeness of cited records; first measure fragments with the selected model, then compare against 1500 and 2500 |
| Minimum relevance score | Hypothetical starting point of 0.3 | First label relevant and irrelevant fragments in the sample, then compare whether irrelevant fragments are filtered at 0.2, 0.3, and 0.4 |
| Question optimization | Enabled | If multi-turn follow-up is tested separately, first supply this order context, then compare whether the subject and order are preserved |
| Result reranking | ReRank enabled | When a reranking model is available, compare ordering against manual labels; the current candidate count has not been measured |
Three common errors
- Treating the minimum relevance score as a permission or entity-isolation control. The minimum relevance score is a filter threshold and does not exclude unrelated businesses or unauthorized records. Check: input the sample question, compare whether returned references still contain records from other stores, and observe whether adjusting the threshold changes that phenomenon.
- Describing RRF as a switch that can be turned on. RRF is the result-fusion process after hybrid retrieval, not an independent configuration item. Check: inspect whether the configuration interface contains a switch named RRF, record the fusion process as documented, and do not invent an independent field on that basis.
- Treating an empty reference output as retrieval failure. Reference content is output as an array, its length may be zero, and the output path still proceeds. Check: use a paper input of an empty array to verify whether the downstream node still executes, record the actual observation, and do not infer a fictional rule.
How to verify the setup
Hypothetical drill sample: the source material reads, "Order number A-2024-001; the buyer applied for a return and refund on March 5, 2024; customer service approved it on March 6; the return tracking number is SF123456; it was signed for on March 9; the refund was completed on March 10. The buyer left a negative review on March 11 with the comment 'slow refund.'" The draft sentence under review reads, "The refund for this order was completed, and the negative review reason was a slow refund." The handover request awaiting judgment reads, "Please confirm from the negative-review handling records whether this order falls within the rectification ledger scope."
- Input the source material and the draft sentence, and check the refund status and the buyer's review text separately. The source states that the refund was completed on March 10 and that the review on March 11 said "slow refund," which supports "refund completed" and the buyer's stated review content. The two dates belong to different events, so they are not required to be equal, and no objective handling duration is inferred from the review.
- Input the handover request and compare whether the negative-review handling records state an authorization or status placing this order in the rectification ledger. If the records contain only the review content and no rectification-attribution field, record the evidence gap and do not infer that the order was or was not included.
- Input the order number from the source material and run full-text retrieval and semantic retrieval in the associated same-vector-model base, then compare whether the returned references contain that number and the corresponding status. If the returned reference is an empty array, check whether the downstream node still executes and record the actual observation, without asserting a fictional rule that retrieval must return empty.