What the documents look like
A general trade export declaration produces a defined set of materials per shipment: the export declaration form, the commercial invoice, the packing list and the contract. These come from China International Trade Single Window, the customs code library, forwarder bills of lading and the company ERP. Compared with general trade import declarations, the export side also touches export tax rebate voucher matching. When something goes wrong, the consequences named for this category are document rejection and inspection, declaration errors, and blocked rebate processing.
For operating reviews and exception retrospectives, the work is field-level comparison across those materials. Retrieval and recall in the knowledge base has a narrower job: locate historical documents, classification records and exception notes so a reviewer can check them by hand. The retrieval step does not decide whether a declaration was correct, whether a rebate was granted, or whether a person is allowed to see a record.
Materials update per shipment. That means the searched library keeps growing as documents are archived. The reference cap is measured in tokens, so the number of items that fit inside one cap depends on how long the retrieved passages are. The returned reference array length is the observable quantity, not a fixed count of documents.
Constraints at the retrieval step
Semantic retrieval works from vector distance. Full-text retrieval works from keyword hits. Precise values such as a goods code, a bill of lading number or a declaration number are better served by full-text or hybrid retrieval than by semantic similarity alone. The minimum relevance value is a 0–1 threshold that applies only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final sort order, and it is not an entity filter or a permission filter.
Problem optimization performs coreference resolution and question expansion before retrieval and adds an extra model call. Reranking requires a rerank model and runs when enabled in the configuration. Image retrieval requires image input plus the relevant model capability. RRF is the fusion process that combines multiple retrieval paths during hybrid retrieval; it is not a switch that an operator turns on. The optional paths drawn in the overall flow are not all executed on every request.
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries on different vector models cannot be associated together. This is checked by comparing the vector model identifier and excluding those that do not match; it does not extend to any external system.
Retrieval does not perform authorization. An empty result still produces an output reference array with length 0, and that downstream path still runs. Forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation.
Configuration and implementation checks
The following is a hypothetical paper exercise, not a record of a real shipment or a completed test. A document clerk at an export trading company is reviewing one general trade export job on 10 March 2025. The three source sentences are: "Export declaration form, declaration date 5 March 2025, goods code 8504409999, destination country Germany." "Commercial invoice amount USD 42000, trade term FOB Shanghai." "Packing list states 12 cartons, gross weight 260 kilograms." The draft sentence to be judged is: "This shipment completed export tax rebate filing on 6 March 2025." The materials needed for the check must sit inside an accessible target retrieval scope or in an explicit manual input.
| Configuration item | Hypothetical starting point for this exercise | How to check |
|---|---|---|
| Associated knowledge bases | Select the export document library that uses the same vector model | Compare vector model identifiers and exclude non-matching libraries |
| Search mode | Run the sample question first with hybrid retrieval | Compare item by item against semantic and full-text results |
| Reference cap | Set to 2000 tokens for this exercise, chosen so the three source passages can be compared for completeness | Compare against 1000 and 4000; actual passage length has not been measured |
| Minimum relevance | Set to 0.4 for this exercise, chosen as a midpoint between a looser and a tighter threshold | Label relevant and irrelevant samples first, then compare against 0.3 and 0.5 |
| Problem optimization | The sample question contains the reference "this shipment", so try it enabled | Compare the original question with the optimized question and check the subject was not altered |
| Result reranking | Try enabled when candidates are numerous | Check the rerank score range and the filtering effect |
Three common errors
- Treating a retrieval hit as completed authorization. The draft sentence claims rebate filing was completed. Retrieval can only return document content. Rebate filing is a business action that needs a separate system implementation. A pending review is not an executed action.
- Treating minimum relevance as an entity or permission filter. The value is a 0–1 similarity threshold. It does not exclude unrelated business or unauthorized records, and retrieved results still require manual checking of subject and period.
- Describing RRF as a switch that can be turned on. RRF is the fusion process for multiple result paths under hybrid retrieval. Reranking and image retrieval each require a model and run when enabled. The optional paths listed in the documented flow are not all executed on every request.
How to verify the setup
- Input the three source sentences and the draft sentence, then search for the original text stating rebate filing completion status and date. The declaration date, invoice amount and carton count each have a record. Rebate filing completion and the 6 March completion date are not stated, so each is recorded as unsupported. Do not substitute another field for completion evidence.
- Input the sample question and run one retrieval. Compare whether the returned references contain the source text above, and observe whether the reference length is 0 or truncated. For the empty-result branch, use the explicit paper input [] to check the downstream design. Do not claim that a hypothetical rule necessarily returns [].
- Compare the vector model identifiers of the associated knowledge bases and exclude those on different vector models. Record the recall items actually observed. Do not run a cross-model retrieval to compare missed recalls, and do not infer what would happen if non-matching libraries were associated together.
Across all three checks, the subject, period and field must match. Only mutually exclusive values in the same field constitute a contradiction. An unstated value, or evidence that covers only a related stage, is recorded as unsupported or pending verification, not as a negation or a conflict. A pending review does not show that operating authority is absent. Signing, reaching an agreement and a contract taking effect are three separate facts. Do not add authorization or effective-date clauses to obtain a definite answer; keep the evidence gap open. Materials needed for the check must sit inside an accessible target retrieval scope or an explicit manual input, and any missing item is recorded as a gap.