What the documents look like
Compliance decisions for platform listings draw on four document families. The product title table records the product name and the wording used for selling points. Detail page assets carry the promotional images and copy. The prohibited advertising terms list names expressions that cannot be used. Inspection reports supply the physical parameter basis. These materials appear together at launch and are retrieved again during weekly review.
The procurement and supplier materials stage adds supplier qualification files, quotation sheets and physical parameters into the same retrieval scope, so that promotional wording can be checked against supporting material. When material is missing, an empty retrieval result does not mean a violation exists. It means the current scope holds no citable basis. The retrieval action itself changes no authorization state of any material and does not confirm supplier qualification.
Two features of this category shape the task. Promotional authenticity and qualification display are handled here, while parcel delivery is not. The most time-consuming steps are checking extreme wording, comparing physical parameters, and replacing non-compliant assets.
Constraints at the retrieval step
Linked knowledge bases must use the same vector model. Knowledge bases with different vector models cannot be linked together. Before retrieval, check the vector model identifier of each base, exclude those that do not match, and record what was observed. Do not run a cross-model retrieval to compare missed recalls.
Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges the two paths through RRF, and ReRank performs a second ordering of candidate text. The minimum relevance value is a filter threshold and takes effect only in semantic retrieval or when result reordering is used. It does not perform permission or entity isolation. The citation limit is measured in tokens and controls how much content each search cites. When question optimization is enabled by configuration, it adds one model call for coreference resolution and question expansion. Image retrieval requires image input and the corresponding model capability, and reordering requires a model and must be enabled by configuration. These paths are not fixed steps of every request. The node outputs an array whose length may be zero, and the empty-result path still completes.
Configuration and implementation checks
The configuration below is built around the hypothetical drill sample on this page. Sample material text: Supplier A provided an inspection report in March 2025 for a certain model of electric kettle. The report states a rated power of 1,800 watts, carries report number QC-2025-0312, and its valid scope is that batch. The draft sentence to be judged is: "This electric kettle has a rated power of 2,000 watts and comes with an inspection report." The draft sentence and the report text hold different values in the rated power field, which is a mutually exclusive value in the same field.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the same-vector-model base containing the inspection report | Check the vector model identifier first, exclude bases with a different model |
| Search mode | Hybrid retrieval | The sample supplies a report number; the draft sentence has no specific model; compare actual recall using the number |
| Citation limit | Set to a token amount covering the full report text | Compare with the actual report length, then adjust |
| Minimum relevance | Set to a medium threshold starting point | Compare using labelled relevant and irrelevant samples |
| Question optimization | Off for this trial | The draft sentence has a complete subject; no reference needs resolution |
| Result reordering | Not enabled for this trial | Few candidates; observe unreordered results first |
These values are hypothetical trial starting points and are not official default values. Only a report excerpt is available; the full text length is not provided. After the full text is obtained, measure it with the selected model, then compare completeness under the citation limit.
Three common errors
- Treating a retrieval hit as confirmed qualification. A returned inspection report fragment only means the fragment was recalled within the retrieval scope. Whether the supplier qualification is valid and whether the report covers the current batch requires manual checking of the report number and valid scope. Check: input the draft sentence and observe whether the returned fragment contains the report number and valid scope fields. If the fragment lacks these fields, record it as pending verification.
- Treating minimum relevance as a permission filter. Minimum relevance only filters low-relevance results and does not distinguish suppliers or business entities. Check: retrieve with the same question and observe whether returned content mixes across suppliers. If it does, record it as a retrieval scope issue and do not judge it as a permission failure.
- Treating RRF as a switch that can be turned on. RRF is the fusion process of multiple result paths in hybrid retrieval, not an independent configuration item. Check: examine whether a switch named RRF exists in the configuration interface. If it does not exist, record it as a difference between the document description and the interface, and do not infer that fusion was not executed.
How to verify the setup
- Input the draft sentence "This electric kettle has a rated power of 2,000 watts and comes with an inspection report," and compare the rated power value in the returned fragment with the report text value of 1,800 watts. If the returned fragment contains the report text with a value of 1,800 watts, the draft sentence and the original text hold different values in the rated power field, and the draft is judged as pending correction. If the returned fragment is an empty array, record it as not recalled in the current retrieval scope and do not judge the draft sentence as correct.
- Input the name of Supplier A and report number QC-2025-0312, and compare whether the returned fragment contains that number and the valid scope. If the returned fragment contains the number but not the valid scope, record it as partial evidence and keep the valid scope gap. If the returned fragment is empty, record it as not recalled and do not infer that the report does not exist.
- Construct a paper input of an empty array and check whether the downstream design still completes the output path when the cited content length is zero. Observe whether the downstream node receives the empty array and continues execution. If the downstream stops, record it as a path design issue. If the downstream continues, record it as empty-result handling working. This check does not claim that a fictional rule retrieval will necessarily return an empty array.
These verifications use the hypothetical sample as input and were not actually executed. Retrieval results do not constitute authorization, approval or qualification confirmation. The validity of supplier material requires separate manual checking.