What the documents look like
Cross-border product compliance and labelling work happens before listing and before import. The evidence base has to support every claim that appears on a product page or in marketing copy. The main material sources are platform policy centres, testing laboratory reports, destination-country regulation libraries, and supplier documents. The documents handled day to day are product test reports, English-language labels, MSDS files, and platform prohibited-item lists.
These materials share a common shape. A single test report covers a specific model, a specific batch, and specific test items. Label wording is constrained by the destination country's language and format requirements. Prohibited-item lists are updated at irregular intervals. The copy drafting task is to place each proposed promotional sentence against these materials, one sentence at a time, and determine whether the existing material supports that sentence.
Constraints at the retrieval step
A knowledge base search node can be linked to one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection is only a way of grouping data for management and does not affect search results. Quoted content is output as an array, and its length can be zero. Even when nothing is found, the output path still runs through.
Search results depend on the search mode. Semantic search relies on vector distance and suits natural-language questions. Full-text search relies on keyword hits and suits numbers, model designations, and proper nouns. Hybrid search uses both and merges them through RRF. The minimum relevance value takes effect only under semantic search or when result reranking is used; it is a filtering threshold, not a permission or entity-isolation mechanism. Retrieval itself does not constitute authorization. Question optimisation performs coreference resolution and question expansion and adds one model call before retrieval.
Hypothetical drill sample: a batch of portable chargers is being prepared for listing. The source material reads: "Test report number TR-2024-0871, client Xingqiao Trading Co., Ltd., sample model PB-20, test item battery overcharge protection, conclusion meets the testing requirements provided by the client, report date 12 March 2024." A separate English label draft states: "Model: PB-20, Input: 5V 2A." The sentence under review is: "This product holds international safety certification and is suitable for sale in all countries." The check compares the subject, object, and validity scope of the draft sentence against the source text. The material supports only a specific model and a specific test item; it states no international certification and no all-country applicability.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the knowledge base holding this batch's test report and label draft | Check each base's vector model identifier first, exclude any that do not match, and do not run cross-model retrieval |
| Search mode | Start with full-text search for the model and report number | Model and number are exact queries; compare hits against semantic search results item by item |
| Reference cap | Start at a value that fully covers one report conclusion paragraph | Test with materials of different lengths to see whether content is truncated, then adjust the cap |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold | This value takes effect only under semantic search or reranking; observe whether irrelevant fragments are filtered out |
| Question optimisation | Enable for follow-up questions containing references; disable for standalone questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous and a rerank model is available | Rerank scores can be used for filtering; compare ordering changes against unreranked results |
Three common errors
- Treating a retrieval hit as authorization. Input: the draft sentence "This product holds international safety certification." Retrieval returns a test report fragment. The comparison is between the report's conclusion field and the draft sentence's claim field. When subject and scope do not match, the sentence is unsupported. Retrieval is recall, not permission or entity isolation; a hit does not establish that the sentence may be published.
- Treating pending review as approved. Input: a label draft marked "pending compliance specialist review." The comparison is between the draft's status field and the value "approved." When the status is pending review, only a pending-verification note can be recorded; the label cannot be written as finalised.
- Forcing knowledge bases with different vector models into one link. Input: the vector model identifiers of two knowledge bases. Compare whether the two identifiers match. When they differ, exclude the non-matching base and record the actual observation. Do not proceed to infer the runtime behaviour of a combined link.
How to verify the setup
- Use the sample draft sentence "This product holds international safety certification and is suitable for sale in all countries" as input, and search within the linked knowledge base. Compare the model, test item, and conclusion fields of the returned references. The expected phenomenon is that the quoted content contains no statement of international certification or all-country applicability, and the sentence is judged unsupported.
- Use the report number "TR-2024-0871" as input and run full-text search. Compare whether the returned content contains the report text corresponding to that number. The expected phenomenon is that the number is hit and the quoted content can be traced to that report. If no hit occurs, record it as a retrieval gap.
- Use an empty array as a paper input to check the downstream design. Compare whether the downstream node still runs the output path when the reference length is zero. The expected phenomenon is that the path continues to execute and produces no quoted content. Do not claim that a fictional rule-based search necessarily returns an empty result.