What the documents look like
General trade export declaration work moves on four recurring documents: the export declaration form, the commercial invoice, the packing list and the contract. These come from the China International Trade Single Window, the customs commodity code library, freight forwarder bills of lading and the company ERP. Each shipment refreshes them, so the knowledge base holds per-shipment material whose field values differ across documents for the same transaction. Compared with general trade import declaration, export declaration also affects export tax refund voucher matching. The stated consequences of getting it wrong are document return for inspection, declaration error and blocked tax refund. For retrieval and recall, this means a question about one shipment may need field-level comparison across several documents rather than a single answer lookup.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together. The smallest search unit is the knowledge base; a collection only classifies data and does not affect search results. A single retrieval may combine question optimization, semantic retrieval, full-text retrieval, hybrid retrieval, image retrieval and rerank. These paths trigger according to input and settings, and not every request runs all of them. Image retrieval requires an image input and the relevant model capability; rerank requires a rerank model and must be enabled. The citation limit is measured in tokens. The minimum relevance score takes effect only under semantic retrieval or when rerank is used. Semantic recall and relevance scores do not perform permission or entity isolation, and retrieval itself is not authorization.
Assumed paper exercise: a foreign trade company's documentation clerk prepares to hand over one export transaction to a customs broker. The source material reads: "Export declaration number HYP-2024-0871, commercial invoice amount USD 42,600, packing list showing 12 cartons, contract signing date 5 March 2024, buyer overseas Company A." A proposed sentence reads: "This shipment has passed customs inspection and tax refund voucher matching is complete." The sentence asserts a subject, an object and a status, while the source material supplies only document fields and states no inspection conclusion or refund matching result. The assertion is unsupported.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison object |
|---|---|---|
| Linked knowledge bases | Link only export declaration material bases using the same vector model | Compare against the node condition requiring the same vector model |
| Search mode | Hybrid retrieval combining semantic and full-text retrieval | Compare against single-mode behavior on code-type and natural-language questions |
| Citation limit | Set an assumed starting token amount that still allows longer documents to be cited in full | Compare against actual document fragments of differing length |
| Minimum relevance | Label relevant and irrelevant samples first, then take an assumed starting value | Adjust after comparing against the labeled results |
| Question optimization | Enable for multi-turn follow-ups containing references | Compare subject consistency between the original question and the optimized question |
| Rerank | Enable when candidates are numerous | Compare against candidate order without rerank |
Three common errors
- Treating a retrieval hit as authorized execution. The proposed sentence claims inspection passed. Retrieving declaration form fields does not constitute an inspection conclusion and does not constitute completed refund matching.
- Forcing a link across vector models. The node requires linked knowledge bases to share the same vector model. When identifiers differ, exclude the mismatched base first; do not run a cross-model retrieval to compare missed recalls.
- Confusing the citation limit with a thinking budget. The citation limit controls the token volume of cited content. A thinking budget is a separate concept. The two are not interchangeable.
How to verify the setup
- Input the assumed source material and the proposed sentence. Compare the sentence's assertions against the fields stated in the source. Observe whether only document fields appear with no inspection or refund conclusion. An unsupported assertion is the expected comparison result.
- Input a multi-turn follow-up containing a reference. Compare the original question with the retrieval terms after question optimization. Observe whether the subject has been altered. Consistent subjects satisfy the condition.
- Input an explicit empty citation array []. Check whether the downstream chain still proceeds. A zero-length output with the chain continuing is documented existing behavior. Do not claim that a fictional rule necessarily returns [].
This page addresses only general trade export declaration / order fulfillment and after-sales question answering / knowledge base retrieval and recall. Retrieval or a form is not authorization to execute.