What the documents look like
Import and export permit handling takes place before customs declaration. Its outcome decides whether a shipment may lawfully cross the border, which separates this category from import declaration work that happens later in the chain. The procurement and supplier-material step works with automatic import licences, dual-use item licences, contracts and product descriptions. These materials arrive from commerce authority systems, customs commodity codes, supplier files and customer end-use statements. Updates are irregular and occur per shipment.
Fields that need precise recall include licence numbers, commodity codes, supplier names, contract clauses and end-use descriptions. The vector model and index step exists so these materials can be retrieved correctly inside a knowledge base for later manual verification. Retrieval supports review; it does not replace the review.
A hypothetical drill sample is used throughout this page. A trade compliance officer receives a copy of an Automatic Import Licence. The licence states that the applicant is "Hualing Machinery Manufacturing Co., Ltd.", the commodity is "CNC lathe", the licence number is "AUTO-2025-0731", and validity runs to "31 December 2025". An attached draft contract states that the supplier is "Hualing Machinery Manufacturing Co., Ltd." and the delivery date is "November 2025". A draft sentence awaiting review reads: "This licence already covers the current purchase, so shipment can be arranged." Review is limited to the licence face and the draft contract text above.
Constraints at the vector indexes step
Knowledge bases linked for retrieval must use the same vector model. Before linking, check the actual model identifier and the supported input type of each knowledge base. One source data block can be associated with several indexes, and the index content and the source text are checked separately. Index content should be complete and focused; retrieval coverage and index length are weighed against the sample at hand.
Image retrieval depends on whether image descriptions or image vectors were generated at ingestion time, not only on what is submitted at query time. Semantic recall and relevance scores do not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion step inside hybrid retrieval, not a switch that is turned on. Reranking requires an available model and must be enabled according to its settings. The empty-result output path still completes, and the reference array length can be 0.
Configuration and implementation checks
The table lists only checks relevant to the sample on this page. Values are hypothetical starting points for this drill, not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Knowledge base holding the licence and contract | Check the vector model identifier of each base first and exclude any that differ |
| Search mode | Hybrid retrieval | Licence number and commodity code are exact fields and run alongside semantic recall |
| Minimum relevance | 0.4 | Hypothetical starting point; adjust after comparing cited source text on the same question set |
| Reference limit | 3000 tokens | Hypothetical starting point; compare against full coverage of the licence face fields |
| Question optimisation | Left off for now | The sample question has a complete subject and predicate; enabling it adds one model call |
| Rerank | Not enabled for now | The candidate set is small; observe fused results before deciding |
Three common errors
- Linking knowledge bases that use different vector models. Input: the model identifiers of two knowledge bases. Compare whether the two identifiers match. Observe whether the linking action is permitted; when they differ, record the mismatch and exclude the base, and do not run cross-model retrieval.
- Treating minimum relevance as a permission or entity filter. Input: a query containing "Hualing Machinery Manufacturing Co., Ltd.". Compare whether returned references may still include material about other entities. Observe that relevance affects ordering and filtering only; permission isolation requires a separate implementation.
- Treating the draft sentence "This licence already covers the current purchase, so shipment can be arranged" as a verified conclusion. Input: the licence validity end date and the contract delivery date. Compare the two date ranges and observe whether the licence states that it covers the goods under this contract. When it does not state this, record the evidence gap and do not supply an authorisation clause.
How to verify the setup
- Input the sample licence face text and the draft contract text. Compare whether the licence number, the entity name and the commodity name all appear in the reference content. Observe whether the reference array contains the corresponding source passages; when it does not, record the field as not recalled.
- Input the draft sentence. Compare its "already covers" wording against the licence validity end date and the contract delivery date. Observe whether the licence face states that the goods under this contract fall within the validity period; when it does not state this, the sentence remains pending verification.
- Input a licence number that is clearly not in the knowledge base. Compare the returned reference array. Observe whether the downstream path still completes when the array length is 0; do not conclude from this that any fictional rule always returns an empty array.
Across all three checks, retrieval and relevance scores do not establish authorisation. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where a field is not stated in the supplied material, that field remains unknown.