What the documents look like
General trade import declaration work centres on the import declaration form, commercial invoice, packing list, bill of lading and contract. Source material also includes the International Trade Single Window, customs tariff schedules, freight forwarder arrival notices and foreign exchange payment documents. Compared with export declaration, import declaration places weight on dutiable value, freight and insurance, and tax payment. Those fields sit across several documents, and the same shipment is held at different stages by the importer, the customs broker and the customs manager. Data updates per shipment. Knowledge base retrieval must cover question-and-answer content and clause fragments drawn from these documents.
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 using different vector models cannot be linked together; the identifiers must be checked first and non-matching bases excluded. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges semantic and full-text results through RRF, which is a fusion process rather than a switch that can be turned on. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and its 0 to 1 value filters low-relevance results. Question optimisation performs coreference removal and question expansion, and adds one model call before retrieval. The retrieval step returns reference content only, as an array whose length may be 0, and does not itself perform permission judgement or entity isolation.
Hypothetical paper exercise sample: a customs manager at an importer intends to place a handover request into the knowledge base retrieval scope. The source text for checking reads: "Bill of lading number HBL-2024-0317, consignee Company A, goods are valve fittings, arrival date 17 March 2024. Commercial invoice INV-2024-0317 states the transaction term CIF, and freight and insurance are already included in the invoice total. Contract CT-2024-0317 sets the payment condition as 30 days after sight of the bill of lading copy." The draft sentence to be assessed reads: "Freight and insurance for this shipment need no separate declaration, and the invoice total may be cited directly at handover." The authorisation boundary for this request is: the customs manager submits retrieval and handover records only, and this does not include tax declaration approval or outward payment execution.
Configuration and implementation checks
| Configuration item | Hypothetical starting point on this page | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only import declaration document bases that use the same vector model | Check that the base identifier and the vector model identifier match, and exclude those that do not |
| Search mode | First test the same question separately under semantic retrieval and full-text retrieval | Compare the reference content returned by the two modes, then decide whether to use hybrid retrieval |
| Reference cap | Set a hypothetical trial starting point of 3000 tokens for this exercise | Test with bill of lading and invoice fragments of different lengths to see whether content is complete, and compare against a shorter cap |
| Minimum relevance | Set a hypothetical trial starting point of 0.5 for this exercise | Label relevant and irrelevant samples first, then compare whether filtered reference content still contains irrelevant fragments |
| Question optimisation | Enable for follow-up questions such as "what is the second point", disable for complete document questions | Compare the original question with the optimised question, and check whether the subject has been altered |
| Result reranking | Enable when a reranking model is available and candidates are numerous | Compare scores and ordering before and after reranking, and confirm whether it runs as configured |
Three common errors
- Treating reference content returned by retrieval as a permission credential. Input is the handover request and the retrieved result array; compare whether the retrieval output contains an authorisation field; the observable result is that reference content holds text fragments only and does not constitute approval or an execution basis.
- Reading minimum relevance as an entity or permission filter. Input is a query containing an unrelated company name; compare whether returned results are excluded by the relevance threshold; the observable result is that the threshold affects similarity scores only and does not exclude unauthorised records.
- Linking knowledge bases that use different vector models. Input is two knowledge bases whose vector model identifiers differ; compare the base identifiers with the model identifiers; the observable result is that identifiers which do not match must not be linked together, and only the actual observation is recorded.
How to verify the setup
- Take the draft sentence from the hypothetical sample, "Freight and insurance for this shipment need no separate declaration, and the invoice total may be cited directly at handover", and place it in the retrieval scope. Compare the transaction term and the freight and insurance wording stated on invoice INV-2024-0317. If the reference content returns only the invoice total field and does not state authorisation for a separate freight and insurance declaration, the draft is judged unsupported and the evidence gap is recorded.
- Take bill of lading number HBL-2024-0317 and arrival date 17 March 2024, and run a full-text retrieval. Compare whether the returned fragment contains that number and date. If the return is an empty array, carry the empty result forward to check the downstream design, without asserting that an invented rule must return empty.
- Take the authorisation boundary in the handover request, "submit retrieval and handover records only", and run retrieval. Compare whether the output contains a tax declaration approval field or a payment execution field. If the output contains reference content only and no approval field, record that retrieval does not carry authorisation execution, and keep the item to be confirmed as unsupported.