What the documents look like
Compliance and labelling evidence for cross-border goods appears before listing and before import, and it centres on product safety and label content. In the inventory procurement and supplier material stage, the documents handled include product test reports, English labels, MSDS files, and platform prohibited-item lists. Sources span platform policy centres, testing laboratory reports, destination-country regulation libraries, and supplier material. These materials are revised on no fixed schedule; each new product introduction can bring a new version.
One test report may correspond to requirements in several destination countries. The version of an English label and the physical packaging need item-by-item comparison. The revision date of an MSDS and the batch information supplied by the supplier need to be aligned. Retrieval works across these materials, which vary widely in length and format. The recalled output needs to preserve subject, object, status, and validity scope so that later human review can use it.
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. Search results are output as an array of references, and the length can be 0, which means the output path still runs even when no content is found. Semantic retrieval relies on vector distance; full-text retrieval relies on keyword matches; hybrid retrieval merges the two paths through RRF; ReRank performs a second ordering of candidate text and returns a score from 0 to 1. The minimum relevance value takes effect only in semantic retrieval or when result re-ranking is used, and it is a filter threshold rather than a final ordering rule. The reference cap is measured in tokens and does not use top k. When question optimisation is enabled, it adds one model call before retrieval for coreference resolution and question expansion.
Retrieval itself does not carry out authorisation decisions. A similarity score does not distinguish business entities and does not filter unauthorised records. Knowledge bases that use different vector models cannot be linked together; the vector model identifier needs to be checked first, non-matching bases excluded, and the actual observation recorded. Cross-model retrieval is not run in order to compare missed recalls.
Hypothetical exercise sample: a supplier material library holds one English label draft. The source text reads: "Product: Stainless Steel Water Bottle. Batch: B-2401. Country of destination: Germany. Label language: English. Revision date: 2024-03-15." An inventory procurement staff member intends to submit a handover request: "Please confirm that this batch label can be used directly for listing on the German market, with no additional German-language marking required." The subject in this request is the supplier and batch B-2401, the object is the English label draft, the status is draft, and the validity scope is the German market. The source text contains no record of a German-language marking requirement and no platform review conclusion.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical exercise | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Select one library holding product test reports and English labels | This library and the supplier material library use the same vector model; check the model identifier before linking |
| Search mode | First run the same question through semantic retrieval and full-text retrieval separately | Compare whether the two recall sets cover the batch number and destination-country fields |
| Reference cap | Set a provisional token value that covers the full text of a single label draft | Test with materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Set a provisional mid-range value | Mark relevant and irrelevant samples first, then compare filtered results |
| Question optimisation | Leave off initially | The current question already contains the subject and batch number; compare whether the subject is altered when the option is on |
| Result re-ranking | Leave off initially | When the candidate set is small, compare ordering before and after re-ranking |
These values are starting points for this hypothetical exercise and are not official default values. Each row in the configuration table corresponds to a checkable setting or a manual verification action, and does not include database internal implementation details.
Three common errors
- Treating a similarity score as a permission filter. Minimum relevance only filters low-relevance results; it does not distinguish business entities and does not exclude unauthorised records. The appearance of a record in retrieval output does not mean the current operator is authorised for that record.
- Treating an intended request as an approved listing. The label is a draft, with no platform review conclusion and no German-market label basis supplied. The claims of direct listing and of no additional marking required each remain unverified. The intended submission is not reworded as pending review, and no label regulation is invented.
- Linking knowledge bases that use different vector models. The vector model identifier needs to be checked before linking, and non-matching bases excluded. The runtime result of a cross-model link is not described in the documentation, so missed recalls or redundant returns cannot be asserted.
How to verify the setup
- Input the English label draft source text and batch number B-2401 from the hypothetical sample, and compare whether the semantic retrieval and full-text retrieval recall sets contain the destination-country field and the revision date. If neither path returns this batch number, record it as not recalled and check whether the data has completed training.
- Input the intended handover request sentence, and compare whether the subject of the retrieval query remains the supplier and batch B-2401 with question optimisation on and off. If the subject is replaced by another batch or another destination country after optimisation, record it as question optimisation altering the subject, turn the option off, and retry.
- Input an explicit paper empty-reference array [] and check whether the downstream design has a branch for handling empty results. If that branch is absent, record it as a design gap. This step only checks the design; whether a real query returns empty and whether the path runs remain untested.