What the documents look like
Import and export permit handling takes place before customs declaration. The result of this stage determines whether a shipment may lawfully cross a border. The materials involved in this task include automatic import licences, dual-use item licences, contracts, and product descriptions. These materials arrive in Chinese, English, and the language of the supplier's country, sometimes within a single shipment file. Vector search carries cross-language understanding, so a Chinese question can match an English knowledge point. This property has direct value for multilingual permit files.
Permit files update per shipment. Licence category, commodity code, and end-use description differ from one shipment to the next, so each retrieval needs a shipment scope. A single retrieval step cannot be treated as covering every shipment in a library.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Before linking, compare the vector model identifier of each candidate library, exclude those that do not match, and record what was observed. Retrieval searches the whole library; a collection is only a classification layer and does not change search behaviour. Reference content is output as an array, and its length may be zero. When nothing is found, the output path still completes.
Semantic retrieval, full-text retrieval, and hybrid retrieval each have their own conditions. Queries that turn on numbers, models, or proper nouns suit full-text or hybrid retrieval. Natural-language questions suit semantic or hybrid retrieval. Hybrid retrieval merges the two result paths through RRF, and similarity filtering cannot be applied directly to the merged set, so a ReRank model is usually needed to reorder candidates and to filter by its score. The minimum relevance value is a filter threshold. It applies only during semantic retrieval or when result reranking is in use. It does not perform permission filtering or entity isolation. A retrieval result does not constitute an authorization. Whether a licence is valid and whether an approval has completed require separate verification.
Hypothetical paper exercise sample: In one shipment file, the product description reads "This product is used for industrial coating equipment, model AC-220, and the export destination is Malaysia." A contract excerpt reads "The buyer shall obtain a dual-use item licence before shipment." The draft sentence under review reads "The shipment has obtained the dual-use item licence and may be arranged for dispatch." The sample contains no record of an issued licence. The draft sentence's licence status is unsupported by the sample.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Two permit libraries that use the same vector model | Compare vector model identifiers first, then link; exclude any library whose identifier does not match |
| Search mode | Hybrid retrieval | The sample contains model AC-220 and a natural-language end-use description; compare semantic and full-text recall results |
| Reference cap | Measured in tokens; starting point set to hold the full product description paragraph | Test with short and long materials to see whether content stays complete, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then tune the threshold | This value applies only during semantic retrieval or result reranking |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when a ReRank model is available | Compare the rerank score with the vector score and filter by the score |
Three common errors
- Treating the licence status in the draft sentence as a verified fact. The sample records only a contract requirement to obtain a licence before shipment; there is no issuance record. Check: input the draft sentence and the contract excerpt, compare how each describes the licence status, and when the draft says "has obtained" while the source has no corresponding record, mark the claim as unsupported.
- Linking knowledge bases that use different vector models. Check: input the vector model identifier of each library, compare whether the identifiers match, exclude any library whose identifier differs, and record the observation. Do not run a cross-model retrieval.
- Treating minimum relevance as permission filtering. Check: input a query that touches a record outside the authorized scope, compare the returned content with the authorized scope, and note that a relevance score does not change a record's authorization status. Authorization requires a separate check.
How to verify the setup
- Input the hypothetical sample's product description and the question "What is the export destination of AC-220?" Compare whether the returned content includes "Malaysia." The condition is met when the returned content contains that destination and the reference array length is greater than zero. When an empty array is returned, record it as a recall miss and check whether the data has finished training.
- Input the draft sentence and the contract excerpt, and compare how each describes the licence status. When the draft's "has obtained" has no corresponding record in the source, mark it as unsupported, keep the evidence gap open, and do not supply a fabricated authorization clause.
- Input a paper empty result
[]and compare how the downstream node is designed to handle that array. The condition is met when the downstream can accept a zero-length reference and continue execution. When it cannot, record it as needing adjustment, and do not claim that a fictional rule-based search is guaranteed to return an empty array.