What the documents look like
Cross-border product compliance and labelling retrieval draws on four document families: product test reports, English label drafts, MSDS files, and platform prohibited-item lists. These documents form an evidence chain before listing and import, with product safety and label evidence at the centre. Updates are irregular; each new product launch can produce new versions, and a correspondence exists between the version of a test report and the version of a label draft. During operations handover, a retrieval result can only show what a document says. It cannot show who is authorised to edit a label or who is authorised to approve a listing. A knowledge base search node can associate one or more knowledge bases that use the same vector model, and retrieval itself carries no authorisation state. Handover records need to distinguish three object types: source material text, a drafted sentence, and a pending request.
Constraints at the retrieval step
The knowledge base search node outputs references as an array, and the length can be 0, meaning the output path still runs even when no content is found. Semantic retrieval, full-text retrieval, and hybrid retrieval are compared separately against test questions; no single mode is assumed to suit every document type. Numbers, model designations, and proper nouns suit full-text or hybrid retrieval; natural-language questions suit semantic or hybrid retrieval. Minimum relevance takes effect only under semantic retrieval or when result reranking is used, and it is a filtering threshold that does not perform permission or entity isolation. The reference limit is measured in tokens, and because chunk lengths vary widely in a mixed knowledge base, tokens control the reference scope. Question optimisation performs coreference resolution and question expansion and adds one model call. Reranking requires a model and must be enabled by configuration, and image retrieval requires image input and model capability. Knowledge bases using different vector models cannot be associated together; verify the identifier first and exclude those that do not match. RRF is the fusion process for hybrid retrieval results, not a switch.
Hypothetical paper exercise sample: a batch of portable power banks is intended for export to a destination country. The source material provided for checking consists of three sentences: "Product test report number TR-2024-0871, issue date 12 March 2024, conclusion: pass." "English label draft states model PB-100, rated capacity 10000mAh, and does not state a UN38.3 test number." "Platform prohibited-item list, April 2024 version, lists lithium battery products without a stated UN38.3 number as pending supplementary material." One handover request to be assessed reads: "Please search the knowledge base to confirm this model has completed compliance approval and can be listed." This request treats a retrieval result as an approval conclusion. Retrieval can only return document content; approval status requires a separate system implementation.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the product test report base and the platform prohibited-item list base that share one vector model | Both bases may be associated only when their vector model identifiers match; verify identifiers first and exclude bases with a different model |
| Search mode | Use full-text retrieval for number-type questions and hybrid retrieval for label-description questions | Compare the references returned by the three modes against the test questions; do not assume one mode for all material |
| Reference limit | Starting value set to a token amount that covers one test report summary | Test with material of differing lengths to see whether content stays complete, then adjust the limit |
| Minimum relevance | Starting value set to a threshold that filters clearly unrelated passages | Label relevant and unrelated samples first, then adjust the threshold; the value takes effect only under semantic retrieval or reranking |
| Question optimisation | Enable for multi-turn follow-up questions, disable for explicit number questions | Compare the original question with the optimised question and check whether the subject was altered incorrectly |
| Result reranking | Enable when candidate results are numerous | Requires a reranking model enabled by configuration; the rerank score can be used for filtering |
Three common errors
- Treating a retrieval reference as an authorisation conclusion. The input is the handover request "Please search the knowledge base to confirm this model has completed compliance approval." The comparison object is the document text returned by retrieval against an approval status field. The retrieval result presents only test report and prohibited-item list content; approval status is not stated in the material, so it is unsupported and cannot be judged as approved.
- Treating pending supplementary material as a prohibition conclusion. The input is the prohibited-item list sentence "lists lithium battery products without a stated UN38.3 number as pending supplementary material." The comparison object is the value "pending supplementary material" against "prohibited from sale." These are different values of the same field; pending supplementary material does not constitute a prohibition, and the evidence gap must be recorded while awaiting a later version.
- Associating knowledge bases that use different vector models. The input is two knowledge bases with different vector model identifiers. The comparison object is the vector model identifier of each base. When identifiers do not match, exclude the non-conforming base, do not run cross-model retrieval, and record the identifier difference actually observed.
How to verify the setup
- Against the test report text in the hypothetical sample, enter the question "What is the conclusion for TR-2024-0871?" and compare whether the returned reference contains that number and the conclusion "pass." The condition is met when the reference content matches the source text. If an empty array is returned, record the empty result and check whether the data has completed training.
- Against the handover request in the hypothetical sample, enter "Please search the knowledge base to confirm this model has completed compliance approval" and compare whether an approval status field appears in the returned references. The condition is met when the references present only document content and no approval conclusion. If a downstream design treats the reference as approval, record that gap and check whether the downstream connects to a separate approval system.
- Against the prohibited-item list text in the hypothetical sample, enter "Is a product without a stated UN38.3 number prohibited from sale?" and compare the value "pending supplementary material" in the returned reference. The condition is met when the reference preserves the source wording and no prohibition determination appears. If the reference is missing, record the evidence gap and check again after an updated list version is available.