What the documents look like
Import declaration work for general trade turns on five document types: the import declaration form, the commercial invoice, the packing list, the bill of lading, and the contract. Compared with export declaration, the emphasis falls on customs value, freight and insurance charges, and tax payment. Confirming the transaction term and completing the freight and insurance fields are the two most time-consuming steps, and each shipment refreshes the whole document set.
These facts shape retrieval for product and marketing copy drafting. The passages a draft needs to cite cluster around transaction terms, how freight and insurance are treated, tax composition, and specific document fields. A search query that contains only a product name will rarely surface the paragraphs that describe price terms, because the wording of those paragraphs is built from clause and field language. Within one knowledge base, a collection serves only as a management grouping; search runs against the whole library. A single data item can map to several vectors, and a hit on any one of them recalls that item.
Constraints at the retrieval step
The retrieval step for copy drafting exists to return cited content from import documents and price-term explanations. It does not verify company identity, does not judge whether a declaration form is authentic, and does not perform tax calculation. A knowledge base search node can associate one or more knowledge bases that share the same vector model; libraries built on different vector models cannot be associated together. That restriction applies only to the libraries associated with this feature. It does not extend to business-registration websites or any other external system.
The following is a hypothetical paper exercise. Three source sentences for comparison: "The transaction term under this contract is CIF, and freight and insurance are already included in the invoice amount." "The commercial invoice states freight and insurance totalling USD 1,200, already included in the unit price." "The customs value declared on the import declaration form is based on the invoice amount; freight and insurance are not added again." The draft sentence to be assessed: "This product is imported on CIF terms, freight and insurance are included in the invoice amount, and the customs value is based on the invoice amount." The exercise checks whether the draft sentence matches the three source sentences above. No real customer, company, or identification number is involved.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | One library containing import document explanations | The library uses one vector model; libraries on other vector models are not associated together |
| Search mode | Hybrid search | Run the same test question through semantic, full-text, and hybrid modes and compare the recalled content |
| Citation limit | 800 tokens | Hypothetical starting point, chosen because the sample passages are short; compare against 1200 tokens for completeness of returned content |
| Minimum relevance | 0.35 | Hypothetical starting point, chosen because relevant and irrelevant samples must first be labelled; compare against 0.5 for the items remaining after filtering |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled | Used when a rerank model is available; compare candidate order before and after reranking |
These values are starting points for this hypothetical exercise, not official defaults. The citation limit is measured in tokens. Minimum relevance takes effect only under semantic search or when result reranking is used. RRF is the fusion step applied after hybrid retrieval and is not a separate switch. Image retrieval requires an image input and model support, and reranking requires a model enabled in the settings; these paths do not indicate that every request executes all of them.
Three common errors
- Treating the draft phrase "freight and insurance are included in the invoice amount" as directly conflicting with the declaration form wording. The declaration form says freight and insurance are not added again, while the invoice says they are included in the unit price; these point in the same direction. The comparison must return to the three source sentences and check each field, rather than drawing a conclusion from a keyword difference.
- Treating minimum relevance as a permission or entity filter. That parameter filters low-relevance results; it cannot exclude unrelated businesses or unauthorised records, and the retrieval step carries no authorisation judgement.
- Associating libraries on different vector models and then inferring missed recalls or redundant results. The correct approach is to check each library's vector model identifier first, exclude those that do not match, and record what is actually observed without inventing run outcomes.
How to verify the setup
- Input the three hypothetical source sentences and the draft sentence, then compare whether the draft aligns with the source on three fields: transaction term, treatment of freight and insurance, and basis of customs value. The expected observation is that each field can be matched to wording in the source; if a field is not stated in the source, record it as unsupported.
- Run the same test question once under semantic search and once under full-text search, then compare the recalled content from each path. The expected observation is that a difference between the two paths is visible, and that it is recorded which path recalled the passage containing the transaction term. A first-position hit or complete recall is not guaranteed.
- Run the search once with the citation limit at 800 tokens and once at 1200 tokens, then compare whether the returned content covers the sample passages completely. The expected observation is that the number of returned items or the truncation point is visible in both runs. If a downstream node receives an empty citation array, handle it as a zero-length output path that still completes, without asserting that retrieval must return empty.
The public technical facts checked for this feature: the knowledge base search node can associate one or more knowledge bases that use the same vector model; search modes should be compared against a test question rather than assuming one mode suits all material; the citation limit is measured in tokens and should be tested with material of differing lengths before adjustment; minimum relevance takes effect only under semantic search or result reranking, and relevant and irrelevant samples should be labelled before tuning the threshold; question optimisation performs coreference resolution and question expansion, so the original and optimised questions should be compared to check whether the subject was altered.