What the documents look like
Permit handling happens before declaration and decides whether a given shipment may lawfully cross a border. For copy drafting in this category, the retrieval corpus centres on automatic import licences, dual-use item licences, contracts and product descriptions. Sources named in the category brief include the commerce authority system, customs commodity codes, supplier material and customer end-use statements. Material is refreshed irregularly and per shipment, and the commodity wording, end-use wording and permit status differ from one shipment to the next.
For drafting work this means a retrieval hit returns candidate source text only. It does not confirm permit status. End-use wording in a product description, the permitted scope and the contract clauses sit in separate documents, and a comparison needs the same subject, the same period and the same field in the original text.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model. Libraries built on different vector models cannot be associated together. Search runs against the whole library; a collection is a management grouping and does not change search behaviour. Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search merges the two paths with RRF, and result reranking applies a ReRank model for a second ordering with a score in the 0-1 range. The minimum relevance value takes effect only under semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation cap is measured in tokens and limits the content length cited per search. Question optimisation performs coreference resolution and question expansion before retrieval and adds one model call. Cited content is returned as an array whose length may be 0, and the downstream chain still runs on an empty result.
Hypothetical drill sample: the source text reads "The goods in this shipment are sealing elements for a certain model of industrial pump, end use is supporting repair of exported equipment, the automatic import licence application has been submitted, status pending review." The draft sentence reads "This batch of sealing elements has obtained the automatic import licence and export declaration can proceed normally." The comparison object is the same shipment, the same period and the licence status field. The source states pending review; the draft sentence states obtained. The two values are mutually exclusive on the same field, and the draft sentence is unsupported. No authorisation clause or effective date is supplied here.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated libraries | Associate one permit material library using the same vector model | Check the vector model identifier of each library first, exclude inconsistent ones, then compare retrieval results |
| Search mode | Full-text search for code-type questions, semantic search for end-use wording | Compare the two result paths against test questions; no single mode is assumed to suit all material |
| Citation cap | Set to 1500 tokens as the starting point for this trial | Compare with actual material of differing lengths to see whether content stays complete |
| Minimum relevance | Set to 0.5 as the starting point for this trial | Label relevant and irrelevant samples first, then compare the filtering effect |
| Question optimisation | Enabled for multi-turn follow-up questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled when candidates are numerous | A rerank model must be available; compare ordering before and after reranking |
Three common errors
- Treating pending review as obtained. Input is the draft sentence and the source text above; compare the licence status field. The source reads pending review and the draft sentence states obtained, so the sentence is unsupported. The check is a field-by-field comparison of source values for the same shipment and the same period.
- Using a similarity score in place of a permit status check. Input is the cited content returned by retrieval and the permit source text; compare whether either states the permitted scope. A high score does not show that a permit has taken effect. The check is whether the cited source text contains the status field.
- Associating libraries across different vector models. Input is the set of libraries to be associated; compare each library's vector model identifier. Identifiers that do not match are excluded. Record the identifiers and exclude the mismatched ones; do not run a cross-model search to compare missed recalls.
How to verify the setup
- Input the draft sentence "This batch of sealing elements has obtained the automatic import licence and export declaration can proceed normally" together with the source text. Compare the licence status field. The source reads pending review and the draft sentence states obtained, so the sentence is judged unsupported and the evidence gap is retained.
- Input a paper empty result array [] and check the downstream chain's design for accepting an empty citation. Confirm that the chain still runs when the citation length is 0. Do not claim that a search under invented rules necessarily returns [].
- Input the identifiers of the libraries to be associated and compare each library's vector model identifier. Only libraries with matching identifiers may be associated together; mismatched ones are excluded and the actual observation is recorded.