What the documents look like
General trade export declaration work generates an export declaration form, a commercial invoice, a packing list and a contract for every shipment. In store service and internal training, these same materials serve two roles at once: they are the business vouchers behind a declaration, and they are the reference material that staff and trainees consult. Export declaration also feeds into export tax rebate voucher matching, so a misquoted field can lead to a returned declaration, a filing error or a stalled rebate. The three slowest steps reported for this category are classifying commodity codes, filling in the declaration form, and checking freight and insurance. Retrieval and recall operate over the library formed by these materials; the returned references are read by store staff and trainees, and retrieval by itself does not change the authorization state of any document.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The node outputs references as an array, and the array length can be 0; when nothing is found, the output path still runs.
Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval merges the two paths through RRF. Rerank reorders candidate text and returns a score from 0 to 1. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires an image input and matching model capability, and reranking requires a model and must be enabled by configuration; these paths do not run on every request. Retrieval and relevance scores do not act as permission or entity isolation.
Hypothetical paper exercise: A document clerk at a foreign trade company submits a set of materials for checking during store training. The source text reads: "The export declaration form for this shipment shows a declaration date of 12 March 2024. The commercial invoice lists a total goods value of USD 48,000. The packing list states 12 cartons. The contract sets the shipment period as April 2024." A drafted sentence reads: "This contract took effect on 1 March 2024, and the tax rebate voucher matching has been completed." The effective date and the rebate matching status in the drafted sentence do not appear in the source text and are unsupported content.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Basis and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the export declaration form library and the commercial invoice library | Both libraries must use the same vector model; check the identifiers before associating |
| Search mode | Start with semantic retrieval | Compare hit differences for the same question against full-text and hybrid retrieval |
| Reference limit | Start at 2000 tokens | Test with real materials of different lengths for content completeness, then adjust |
| Minimum relevance | Start at 0.5 | Label relevant and irrelevant samples first, then tune the threshold; effective only under semantic retrieval or reranking |
| Question optimization | Start with it off | Compare the original question and the optimized question to check whether the subject was altered |
| Result reranking | Not enabled for now | Enable it only with a model present and per configuration, then compare rerank scores |
Three common errors
- Treating a retrieval hit as an authorization voucher. Input the drafted sentence above, compare the returned references against the source text, and observe whether the references contain a contract effective date or a rebate matching status. The source text has no such fields, and a retrieval hit cannot show that the contract took effect or that rebate matching was completed.
- Forcing knowledge bases with different vector models into one association. Input two libraries whose vector model identifiers differ, compare the identifiers, and observe whether an identical identifier exists. Different identifiers exclude the association; do not infer missed recall or redundancy from this.
- Treating minimum relevance as a permission or entity filter. Input a query that touches another company's records, compare the returned references against the query subject, and observe whether the references belong to that subject. A relevance score does not distinguish subject ownership; unauthorized records need downstream implementation controls.
How to verify the setup
- Input the hypothetical source text and drafted sentence above, compare "took effect on 1 March 2024" in the drafted sentence against the date fields in the source text, and observe whether the source text states that effective date. When the source text does not state it, record it as unsupported; do not judge it as a contradiction or a denial.
- Input an exact code query, compare the returned references from semantic retrieval and full-text retrieval, and observe whether full-text retrieval hits the code. If semantic retrieval misses while full-text retrieval hits, record this as a mode difference; a first-position hit is not guaranteed.
- Input an explicit paper empty result [], compare how downstream nodes are designed to handle empty references, and observe whether the output path still runs. The empty array checks downstream handling design only; it does not claim that a real retrieval will return an empty result.