What the documents look like
Marketplace refund and return handling runs continuously by order. A single case can produce a refund application, a return logistics document, inspection photographs, and a refund approval record, with updates occurring per order and daily. In cross-border work these materials commonly contain Chinese, English, and local languages, and one order's application description, logistics milestones, and inspection conclusion may sit in separate documents. Retrieval therefore needs multilingual text, image descriptions, and exact fields such as order numbers inside the recallable range, so that customer service agents and after-sales specialists can check them. Retrieval returns cited content only; responsibility judgment, amount approval, and return-chasing remain with people or downstream systems.
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 using different vector models cannot be associated together; verify the identifier first and exclude those that do not match. Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Semantic retrieval relies on vector distance and suits natural-language questions; full-text retrieval suits exact queries such as numbers and model codes; hybrid retrieval merges both result sets through RRF. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and acts as a filter threshold. Question optimization performs coreference resolution and question expansion and adds one model call. The citation limit is measured in tokens. Image retrieval requires image input and model capability; reranking requires a model and must be enabled by configuration. Optional paths shown in the overall flow diagram do not mean every request executes all of them. Cited content is output as an array whose length can be zero.
Hypothetical paper exercise sample: a refund application reads "Order A-2201, buyer signed on 12 March, submitted a refund application on 15 March, reason: outer packaging damaged, two inspection photographs attached." A return logistics document reads "A-2201 return collected by carrier on 18 March, arrived at the return warehouse on 21 March." The draft sentence to be judged is "The refund application for this order was submitted within seven days of signing and meets the acceptance conditions." Checking this draft sentence requires evidence of the same subject, the same period, and the same field. The source text does not state the specific acceptance rule, so this conclusion remains unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the refund and return knowledge base that uses the same vector model | Check the vector model identifier first, exclude bases with a different model, and avoid cross-model association |
| Search mode | Hybrid semantic and full-text retrieval | Run the same test question through all three modes and compare recall for number-type and natural-language-type material |
| Citation limit | Set a token value that covers the longest inspection description | Test with actual materials of different lengths for content completeness, then adjust |
| Minimum relevance | Start with a middle value, label relevant and irrelevant samples, then adjust | Takes effect only in semantic retrieval or reranking; observe changes in cited content after filtering |
| Question optimization | Enable for multi-turn follow-up scenarios, disable for clear single-turn questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a reranking model is available | Observe reranking scores and citation order; first-position hits are not guaranteed |
Three common errors
- Treating a retrieval hit as authorized execution. The draft sentence "meets the acceptance conditions" is a business judgment. Retrieval returns cited content only; approval and refund actions require downstream implementation, and a retrieval result does not constitute approval.
- Forcing association of knowledge bases with different vector models. The documentation states the node can associate only knowledge bases using the same vector model. Verify the identifier and exclude non-matching bases; do not run cross-model retrieval to compare missed recall.
- Treating minimum relevance as a permission or entity filter. This value filters low-relevance results only. It does not exclude unrelated businesses or unauthorized records, and retrieval does not perform authorization isolation.
How to verify the setup
- Input the refund application text and the draft sentence from the hypothetical sample, then compare whether the draft sentence shares the same subject, the same period, and the same field as the source. When the source does not state the acceptance rule, record the result as unsupported rather than contradictory.
- Input order number A-2201 and run full-text retrieval and semantic retrieval separately. Observe whether the number field is recalled. If the number is not hit, check whether full-text or hybrid retrieval was used, and record the actual observation.
- Use a paper input of an empty array to check the downstream design, confirming the chain still proceeds when the citation length is zero. Do not claim that retrieval for a fictional rule necessarily returns an empty array, and do not guarantee complete recall.