What the documents look like
Cross-border resale returns handle reverse logistics and item condition. This differs from overseas warehouse inbound replenishment, which concerns forward stock positioning. The knowledge base for this page holds return inspection forms, refurbishment records, scrap declarations and relisting checklists. Retrieval serves contract and transaction material comparison by returning citable source passages. These materials update daily or per batch. The retrieval step returns existing content only. It does not decide return grades, schedule refurbishment or rule on scrap.
A paper exercise sample is used below. Source text one: "Return inspection form RMA-2024-0312 states: principal is a Shenzhen cross-border seller; object is 200 units of wireless earphones; inspection status is refurbishable; date scope is the batch of 12 March 2024." Source text two: "Refurbishment record states: the same batch of 200 units has completed exterior cleaning and packaging replacement; refurbishment completion date is 15 March 2024." The draft sentence to be judged: "This batch of 200 units has completed refurbishment and is ready for relisting, and a relisting application may be submitted." The draft does not state whether a relisting checklist has been generated, and does not state whether a relisting application has been authorized. Both points remain unverified.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Reference content is output as an array, and its length can be zero. When nothing is found, the output path still completes. Semantic retrieval, full-text retrieval and hybrid retrieval suit different material. Identifier and model-number lookups suit full-text or hybrid retrieval. Natural-language questions suit semantic or hybrid retrieval. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. The reference cap is measured in tokens and controls how much content a single search may cite. Question optimization performs coreference resolution and question expansion, and adds one model call before retrieval.
Hybrid retrieval combines semantic and full-text recall through RRF. RRF is a fusion step for mixed results, not a switch. Reranking requires a model and must be enabled by configuration. Image retrieval requires image input and matching model capability; it does not run on every request. Bases with different vector models cannot be associated together. Verify the model identifier, exclude any base that does not match, and record what was observed; do not run cross-model retrieval to compare missed recalls.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One base each for return inspection forms, refurbishment records, scrap declarations and relisting checklists | All four must use the same vector model to be associated together; check each model identifier and exclude any mismatch |
| Search mode | Hybrid retrieval | Material contains RMA identifiers and natural-language descriptions; hybrid retrieval uses semantic and full-text recall fused by RRF |
| Reference cap | Hypothetical starting value 3000 tokens | Set by hand for this exercise to observe whether long and short material is cited completely; compare against 1500 tokens and 6000 tokens |
| Minimum relevance | Hypothetical starting value 0.4 | Set by hand for this exercise; takes effect only under semantic retrieval or reranking; label relevant and irrelevant samples first, then compare against 0.3 and 0.5 |
| Question optimization | Enabled for this exercise | Used for coreference resolution and question expansion; compare the original question with the optimized question and check whether the principal was altered |
| Result reranking | Enabled for this exercise | The rerank model reorders candidate text and returns a score from 0 to 1; compare ordering with reranking disabled |
Three common errors
- Treating a retrieval hit as authorization. The draft sentence says a relisting application may be submitted. Retrieval can return the refurbishment record text; it cannot show that a relisting application was approved. Check: input the draft sentence, compare whether the returned references contain an authorization or approval field, and observe that when only a refurbishment completion record appears, the authorization portion is unsupported.
- Associating knowledge bases that use different vector models. If the four bases carry inconsistent vector model identifiers, they cannot be associated together. Check: verify each base model identifier, compare whether all match, and observe that any inconsistent base is excluded and the observation recorded, without running cross-model retrieval to compare missed recalls.
- Writing a pending-review status as completed. If a scrap declaration states only "pending review", scrap cannot be treated as completed. Check: input the scrap declaration source text, compare the status field value, and observe that when only "pending review" appears, the status is unverified and no review conclusion is supplied.
How to verify the setup
- Input the return inspection form source text and the draft sentence from the sample. Compare the draft claim of completed refurbishment against the completion date and quantity in the refurbishment record. Observe that the returned references match the draft on principal, object and quantity, and that no relisting authorization field appears. The result is partially supported, with the authorization portion unverified.
- Input the RMA identifier "RMA-2024-0312". Compare the recall results of full-text retrieval and hybrid retrieval. Observe that the identifier is matched exactly and that the cited content contains the batch corresponding to that identifier. First-position matching and full recall are not guaranteed.
- Input a paper query unrelated to all four bases. Compare the returned reference array. Observe an array of length zero while the downstream path still completes, which checks the empty-result handling design. Do not claim that a fictional rule query always returns an empty array.