What the documents look like
Platform refunds occur continuously per order and are not necessarily tied to a promise made in a single livestream session. Store service and internal training therefore work from materials that arrive order by order: refund applications, return logistics labels, quality inspection photos, and refund approvals. These materials update per order and daily, and they originate in separate systems, including the after-sales backend, customer service chat, WMS receiving, and payment refund records.
For knowledge base retrieval and recall, the form of each material determines the retrieval path. Refund applications and refund approvals are mostly field-based text. Return logistics labels contain tracking numbers and proper nouns. Quality inspection photos are image input, which requires a vision model or a vector model that supports images, together with a corresponding index built at ingestion time. Retrieval word quality and data completeness both affect recall. In store training, questions often omit the subject or the time range, so question optimization can fill in references, and the optimized result still needs human checking that the subject was not altered incorrectly.
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 with different vector models cannot be associated together. When identifiers do not match, the correct action is to check the identifiers and exclude the non-matching base; do not run cross-model retrieval to compare missed recalls. Referenced content is output as an array, and its length can be 0, which means the downstream chain still proceeds even when nothing is retrieved.
The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. That value is a filtering threshold and does not perform permission or entity isolation; a similarity setting does not exclude unrelated business records or unauthorized records. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a switch. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled according to configuration. Optional paths shown in the overall document flow diagram are not executed on every request.
Hypothetical drill sample: a store after-sales specialist receives a refund application. The application text reads: "Order A-1042, buyer signed on March 12, refund application submitted on March 15, reason is item not as described, return tracking number SF-7788 uploaded, two quality inspection photos pending warehouse confirmation, refund approval status is pending review." A separate draft handover sentence reads: "This order has completed refund approval, and the payment has been returned to the buyer." The request to be judged is: include this draft sentence in the store training question-answer base for customer service retrieval and citation.
Configuration and implementation checks
The following items center on the hypothetical drill sample and list only settings documented officially or implementation checks.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the store after-sales base that uses the same vector model | Check the vector model identifier of each base, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Test semantic retrieval first, then test full-text retrieval with the same question | Compare recall of the return tracking number and "item not as described" under the two modes |
| Citation limit | Hypothetical starting point of 2000 tokens for this drill | Test content completeness with refund applications and approval materials of different lengths, then adjust the value |
| Minimum relevance | Hypothetical starting point of 0.5 for this drill | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimization | Enable for follow-up questions such as "what is the second point" | Compare the original question with the optimized question and check whether order number and date were altered incorrectly |
| Result reranking | Enable only when a reranking model is configured | Compare candidate ordering before and after enabling, and confirm the score range and filtering effect |
Three common errors
- Treating a retrieval hit as authorized execution. The input is the draft sentence "This order has completed refund approval, and the payment has been returned to the buyer." Retrieval may recall refund approval materials, but the source text states the refund approval status is "pending review." The retrieval result does not constitute evidence that approval was completed or that a refund was executed.
- Treating a similarity threshold as permission isolation. The input is a store training question. If raising the minimum relevance is assumed to exclude unauthorized records, that confuses a filtering threshold with access control. Retrieval does not perform authorization decisions.
- Forcing knowledge bases with different vector models to be associated together. The input is two knowledge bases with different vector model identifiers. Ignoring the identifiers and associating them directly violates the same-vector-model requirement. The correct action is to check the identifiers and exclude the non-matching base.
How to verify the setup
- Input the hypothetical sample text and the draft sentence. Compare the draft sentence "has completed refund approval, and the payment has been returned to the buyer" with the source text "refund approval status is pending review." Observe whether the draft sentence is judged unsupported. Pending review does not constitute completed approval.
- Input the return tracking number SF-7788 as the retrieval term, and run semantic retrieval and full-text retrieval separately. Compare the recalled content under the two modes and observe whether full-text retrieval matches the tracking number more consistently. Question optimization is not suitable for heavy use on identifier-type queries.
- Input an explicit paper empty result []. Check whether the downstream chain still proceeds, and observe whether the flow continues when the output array length is 0. Do not claim that a fictional rule retrieval necessarily returns an empty array.