What the documents look like
For general trade export declaration work, the daily document set is the export declaration form, the commercial invoice, the packing list and the contract. Data originates from the China International Trade Single Window, the customs code library, freight forwarder bills of lading and the company ERP, and it updates per shipment. In the inventory procurement and supplier materials task, the material to be retrieved typically carries supplier name, commodity code, transaction quantity, unit price, currency, country of origin and trade mode. These fields must agree across the declaration form, the invoice and the packing list; a mismatch in any one of them can lead to a returned declaration, inspection or a filing error. Export declaration also affects the matching of export tax rebate vouchers, so the correspondence between supplier materials and procurement records needs to be recalled accurately at the retrieval stage.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; the vector model identifier must be checked first and non-matching bases excluded. Retrieval across models is not performed in order to compare missed recalls. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, and hybrid retrieval fuses both paths through RRF. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission isolation or entity exclusion. The citation limit 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. Image retrieval requires a corresponding image input and model capability, and reranking requires a model and must be enabled by configuration; these optional paths do not all execute on every request.
Hypothetical exercise sample: a document clerk at a foreign trade company searches the knowledge base for supplier A's procurement record. The material text reads: "Supplier A signed a procurement contract on 10 March 2025, contract number CG-2025-0310, goods are stainless steel fasteners, quantity 5,000 pieces, unit price USD 2.50, trade mode general trade. The contract is valid until 30 September 2025." The draft sentence to be assessed reads: "Supplier A's procurement contract CG-2025-0310 is currently valid and can be used directly for this export declaration filing." The word "currently" in the draft sentence points to February 2026, which is beyond the validity period stated in the contract.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the supplier materials base that uses the same vector model | Check the vector model identifier and exclude bases with a different model |
| Search mode | Hybrid retrieval | Supplier names and contract numbers need full-text hits, while procurement descriptions need semantic matching |
| Citation limit | Hypothetical starting point of 2000 tokens | Test content completeness with actual materials of different lengths, comparing against 500 and 4000 |
| Minimum relevance | Hypothetical starting point of 0.35 | Label relevant and irrelevant samples first, then compare the filtering effect at 0.2, 0.35 and 0.5 |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled | Compare the ordering before and after reranking when there are many candidate results |
Three common errors
- Treating minimum relevance as a permission filter. Minimum relevance is a similarity threshold; it does not exclude unrelated business or unauthorized records. Procurement records of other suppliers may appear in the results, and the subject field must be checked manually. Check method: enter the query for supplier A and observe whether the returned results contain records for supplier B; if they do, the threshold has not performed entity isolation.
- Associating knowledge bases that use different vector models together. Bases with different vector models cannot be associated; the identifier must be checked first and non-matching bases excluded. Check method: check the vector model identifier of each base; if they differ, record the base as not meeting the association condition, and do not run retrieval across models to compare missed recalls.
- Treating a pending-review state as approved. The validity period in the draft sentence has passed and cannot be judged as currently valid. Check method: compare the draft sentence against the date fields in the contract text item by item; the validity runs until 30 September 2025 while the draft sentence points to February 2026, so the judgment is unsupported.
How to verify the setup
- For contract number CG-2025-0310 in the hypothetical sample, enter the search term "CG-2025-0310 supplier A procurement contract" and compare whether the returned content contains the validity field from the contract text. If the returned content contains only the contract number and not the validity period, the citation limit may have truncated a key field, and the setting needs adjustment before comparison again.
- For the draft sentence "Supplier A's procurement contract CG-2025-0310 is currently valid", compare the date reference in the draft sentence against the contract text's validity until 30 September 2025, item by item. If the draft date falls outside the validity period stated in the source, judge it unsupported, record the evidence gap, and do not supply authorization or effective clauses.
- For empty-result handling, use the paper input [] to check the downstream design. If retrieval returns a citation array of length 0, observe whether the downstream path runs through, and record the actual observation without claiming that a fictional rule's retrieval necessarily returns [].