What the documents look like
Cross-border product compliance and labelling material centres on product safety and label evidence. The sources named in the brief 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. Three properties of this material shape retrieval.
First, updates are irregular and every new product listing can introduce new documents. A retrieval library therefore needs collections to separate product lines or destination countries, while the documented search unit remains the whole library: collections classify data for management, and search runs against the library.
Second, test reports and labels carry identifiers, model numbers and chemical substance names. A natural-language question can miss these precise fields, and full-text or hybrid retrieval suits identifier-type queries better.
Third, one source document can map to several vector entries. A test report can carry a summary index and a key-parameter index at the same time, and a hit on any one vector recalls the underlying data.
Constraints at the retrieval step
Chunk lengths differ across test reports and English labels. The citation limit is measured in tokens rather than a top-k item count, which is the documented way to absorb uneven chunk length. The knowledge base search node can link one or more knowledge bases that use the same vector model; bases with different vector models cannot be linked together, so the vector model identifier must be checked and non-matching bases excluded.
Hybrid retrieval merges semantic and full-text recall through RRF. RRF is a fusion step in the retrieval chain, not a separate switch that is turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Retrieval output is an array and its length can be 0, so the empty-result path still completes. Question optimisation performs coreference resolution and question expansion and adds one model call; whether to enable it depends on observed results.
Hypothetical paper exercise
Material text for comparison: "The test report number for this batch of children's water bottles is TEST-2024-0317. The tested item is bisphenol A migration, the result is not detected, and the report date is 17 March 2024. The report corresponds to product model CW-200 and the applicable destination country is Germany."
Draft sentence to be judged: "Children's water bottles of model CW-200 have passed German bisphenol A testing, report number TEST-2024-0317, and can be confirmed to meet German listing requirements."
Judgement point: the material states a test result and a report number. It does not state the conclusion "meets German listing requirements." A listing compliance judgement needs the corresponding clauses from the destination-country regulation library and the platform policy centre, which this material does not provide. The draft sentence is an unsupported conclusion and needs a regulatory basis added or a pending-verification mark.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Select compliance libraries under one vector model | Check the vector model identifier of each library, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Start with hybrid retrieval | Compare full-text recall and semantic recall on identifier questions such as TEST-2024-0317 and observe whether the identifier is hit |
| Citation limit | Set a token amount that covers one test report summary | Test with long and short test report passages and compare whether returned content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust | Takes effect only under semantic retrieval or reranking; observe whether irrelevant fragments are filtered |
| Question optimisation | Enable in multi-turn follow-up scenarios | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous | Requires a rerank model enabled in settings; observe whether ordering changes |
Three common errors
- Treating a retrieval result as an authorisation conclusion. Input "Is TEST-2024-0317 allowed to be listed," and retrieval returns the report text, which does not state a listing permission. The error is equating retrieved evidence with an approval. Check: input that question, compare whether the returned content contains a listing permission field, and mark it pending verification when only the test result is returned.
- Linking knowledge bases across different vector models. Input two libraries with different vector models for linking, expecting broader recall. The error is ignoring the same-vector-model linking requirement. Check: verify each library's vector model identifier, exclude non-matching ones, record the actually linkable library list, and do not run cross-model retrieval to compare missed recall.
- Treating minimum relevance as a permission filter. Input "return only German labels I am authorised to view," raise minimum relevance, and expect unauthorised records to be excluded. The error is that relevance is a similarity threshold and does not perform permission isolation. Check: input that question, compare whether the returned results still contain unauthorised records, and confirm that permission requires a downstream system implementation.
How to verify the setup
- Verification one: identifier recall. Input "What is the test result for TEST-2024-0317" and compare whether the returned content contains the tested item and result for that identifier. If the returned content contains "bisphenol A migration, not detected," then full-text or hybrid retrieval meets the condition for recalling that identifier. If an empty array is returned, record it as not recalled and check whether the data has completed training.
- Verification two: conclusion boundary. Input "Can model CW-200 be confirmed to meet German listing requirements" and compare whether the returned content contains a listing permission conclusion. If the returned content only contains the test report text without a listing permission, judge it unsupported and add a regulatory basis. If the returned content claims listing requirements are met, check whether that conclusion has material support.
- Verification three: empty-result handling. Input an identifier that does not exist in the knowledge base and compare whether the output is an array of length 0. If the output is [], the empty-result path completes and downstream design can handle an empty array. A claim that a fictional-rule query must return [] cannot be made; only the observed output length is recorded.
The above is a paper comparison of a hypothetical sample and was not actually tested. Actual configuration is adjusted according to observed results.