What the documents look like
Permit handling takes place before declaration and determines whether a shipment may lawfully cross the border. The working materials for this category are automatic import licences, dual-use item licences, contracts and product descriptions. Their sources cover the competent commerce authority system, customs commodity codes, supplier material and customer end-use statements. Records are produced per shipment and updated irregularly. Judging which regulatory document applies, preparing the end-use statement and following approval status consume most of the working time.
An operating review or exception review has to answer a narrower question: what state was a given shipment in during a given period, and does a draft statement agree with the source material. Retrieval only locates the relevant source text. It does not confirm authorization, approval outcome or customs clearance. Because updates are irregular, the licence, contract and product description for one shipment can sit at different stages, so retrieval has to align by shipment, by period and by field, and must not treat materials from different stages as one fact.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection only classifies data. Retrieval paths include semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF, a ReRank model reorders candidate text a second time, and the minimum relevance value takes effect only under semantic search or when result reordering is used. The reference limit is measured in tokens. Question optimization performs reference resolution and question expansion and adds an extra model call.
These capabilities serve document location. They do not perform permission decisions. A similarity score is a relevance measure and does not act as an entity or authorization filter. Knowledge bases with different vector models cannot be associated together; the identifier is checked and non-matching bases are excluded, without running a cross-model search to compare missed recalls. The paths drawn in the documented flow do not all execute on every request: image retrieval requires image input and the corresponding model capability, and reordering requires a model and must be enabled by configuration. Empty output is returned as an array whose length may be 0, so downstream design has to accept that result.
Hypothetical paper exercise sample: a foreign trade enterprise has one dual-use item export shipment. The source material reads: "Contract number HX-2024-031, signed 12 March 2024, subject is a device of a stated model, quantity 2 units, buyer is an overseas company; the product description states the final use is industrial inspection; the licence application is still under review and no issuance record is present." The draft sentence to be judged is: "The dual-use item licence for this shipment has been issued and shipment can be arranged." The check compares subject, object, period and status field: the material shows only that review is pending, while the draft sentence claims issuance, so the claim is unsupported. If the draft sentence instead reads "The licence application for this shipment has been submitted and the approval status is to be followed up," it agrees with the source text.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Select one or more bases using the same vector model | If licence, contract and product description sit in separate bases, confirm the vector model identifier matches |
| Search mode | Compare semantic, full-text and hybrid search against the test question | Numbered fields suit full-text or hybrid search; natural-language end-use description suits semantic search |
| Reference limit | Unit is token; test completeness with actual material of differing length | A hypothetical trial starting value is set to cover one licence plus one product description paragraph, compared with a value covering only a short licence sentence |
| Minimum relevance | Takes effect only under semantic search or result reordering | Label relevant and irrelevant samples first, then adjust the threshold; no fixed number is preset |
| Question optimization | Used for reference resolution and question expansion | Compare the original and optimized question across follow-up turns and check whether the subject was altered |
| Result reordering | Requires a ReRank model and must be enabled by configuration | With many candidates, compare ordering before and after reordering; it is not assumed to run on every request |
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is a retrieved licence fragment; the comparison is between the status field in the fragment and the status claimed by the draft sentence; the observable result is whether both share the same subject, period and field. If the fragment shows only that review is pending and the draft sentence states issuance, the claim is unsupported and shipment cannot be arranged on that basis.
- Treating a similarity score as an entity or permission filter. The input is a high-similarity record belonging to another subject; the comparison is between the subject identifier in the record and the target business subject; the observable result is that a record with a different subject does not form evidence for this shipment. Similarity carries no isolation duty, so the identifier is checked separately.
- Associating knowledge bases across different vector models. The input is two bases whose vector model identifiers differ; the comparison is between those identifiers; the observable result is that they cannot be associated together. The actual observation is recorded, without inferring missed recalls or redundancy.
How to verify the setup
- Using the hypothetical sample source text and draft sentence above, input the retrieval question "HX-2024-031 licence status", compare the status field in the returned fragment with the draft sentence, and observe that when the fragment shows only pending review while the draft sentence claims issuance, the claim is judged unsupported. If an empty array is returned, record that the shipment was not matched within the target retrieval scope and keep the evidence gap.
- Input a licence record belonging to another subject, compare the subject identifier in the record with the target business subject, and observe that a record with a different subject does not enter the evidence chain for this shipment. This check verifies the identifier comparison action; it does not verify whether a similarity threshold can exclude unrelated business.
- Input two knowledge bases whose vector model identifiers differ, compare whether the identifiers match, and observe that non-matching identifiers exclude joint association, with the actual observation recorded. This check does not run a cross-model search and does not infer runtime results.
The checks above are paper comparisons. No retrieval, approval or shipment operation was actually executed. A pending review status does not constitute issuance, and the claim in the draft sentence needs support from the source text; a missing issuance record remains missing and is not filled in by inference.