What the documents look like
Returns processing for cross-border resale depends on four document types that record different stages of disposition. The overseas warehouse inspection stage decides the condition of a returned item. The inspection form records appearance and function conclusions. The refurbishment record records part replacement and packaging actions. The scrap form records a determination that an item cannot be resold. The relisting list records entries that may return to sale. These four materials together form the factual basis for drafting product and marketing copy. Materials update daily or per batch, and retrieval for a drafting task uses a specific batch as its scope.
Drafting copy for a returned and refurbished item means confirming the current state of a specific item and the content that state supports. A condition conclusion, a refurbishment action, a scrap determination, and a relisting entry are separate fields at separate stages. Copy that states a fact not carried by these fields exceeds what the materials support.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Search modes include semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. The citation limit is measured in tokens. The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled according to settings. Retrieval results do not perform permission or entity isolation.
A minimum relevance score is a filter threshold, not a final ordering rule. When question optimization, hybrid retrieval, image retrieval, or result reranking is active, final results may come from multiple merged recall paths and do not necessarily follow a single vector similarity order. The citation limit uses tokens rather than a fixed item count because chunk lengths differ across mixed knowledge bases.
Hypothetical paper exercise
Material text for comparison (fictional):
"Batch HX-2405, item is a folding storage box, inspection conclusion is slight scratches on the outer shell and normal function, refurbishment record is outer packaging replacement, disposition conclusion is eligible for relisting."
Draft sentence to be judged: "After refurbishment, the folding storage box in this batch has intact function and can be sold normally."
Comparison targets: whether "intact function" in the draft sentence and "normal function" in the material are the same field with the same value; whether "after refurbishment" matches the action recorded in the refurbishment record; whether the disposition conclusion supports the wording "can be sold normally."
Configuration and implementation checks
| Configuration item | Value or check for this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Associate the bases holding inspection forms, refurbishment records, scrap forms, and relisting lists | The node requires the same vector model; verify identifiers before associating |
| Search mode | Try semantic retrieval and hybrid retrieval separately with the sample question | Identifier-type content suits full-text or hybrid retrieval |
| Citation limit | Hypothetical trial starting point set to a token amount covering one complete inspection record | Adjust after comparing against actual records of differing lengths |
| Minimum relevance | Label relevant and irrelevant samples first, then compare threshold effects | Takes effect only under semantic retrieval or result reranking |
| Question optimization | Compare whether the subject of the original question is altered incorrectly after optimization in multi-turn follow-ups | Used for coreference resolution and question expansion |
| Result reranking | Enable according to settings when candidates are numerous and the question is clear | Rerank scores can be used for filtering |
Three common errors
- Rewriting "normal function" from the material as "intact function," placing an unstated degree judgment into the copy. The check compares the wording of the draft sentence against the wording of the material text; a degree word absent from the material indicates the draft exceeds the supported range.
- Writing "disposition conclusion is eligible for relisting" directly as "can be sold normally," treating a disposition-stage conclusion as a sales fact. The check compares the fact stated in the draft sentence against the field recorded in the material; the two not belonging to the same field means the statement is unsupported.
- Writing a refurbishment record pending review as a completed refurbishment action in the copy. The check compares the record status field; a status of pending review does not constitute a completed action.
How to verify the setup
- Input the draft sentence and the material text from the hypothetical sample, and compare the wording "intact function" against "normal function." The sentence passes only when its wording does not exceed the range of the material text; a degree word absent from the material is recorded as exceeding support.
- Input the handover request "this batch can be sold normally" and compare it against the disposition conclusion field. The request is supported only when the material explicitly carries a sales-permission field; where only a disposition conclusion exists, record it as pending verification and preserve the evidence gap.
- Input a query containing only an image, and compare the returned result against whether the knowledge base has established an image index. The image retrieval path is available only when a valid image index was generated at ingestion and the model supports image input; where no index was established, record that path as unavailable and do not infer text retrieval results.