What the documents look like
Platform product title and detail compliance handles the truthfulness of promotional claims and the display of qualifications. It does not handle parcel delivery. The main materials are product title tables, detail page assets, advertising prohibited-word lists, and third-party inspection reports. Sources include platform rule centers, advertising compliance checklists, brand asset libraries, and inspection reports.
In cross-border work, the title table, detail assets, and inspection report for the same product may exist in Chinese, English, or other languages separately. Retrieval needs to match the same object across languages. Materials update weekly and at each new product launch. At launch, the title table and detail assets change together. Inspection reports update by batch.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be associated together. Before associating, check the vector model identifier of each base, exclude those that do not match, and record the actual observation. Do not run cross-model retrieval to compare missed recalls.
Semantic retrieval supports cross-language understanding. A Chinese question can match an English knowledge point, which suits natural-language phrasing. Full-text retrieval suits exact queries such as numbers, model codes, and proper nouns. Hybrid retrieval uses semantic recall and full-text recall at the same time and merges results through RRF. RRF is a fusion process for hybrid results, not an independent switch that can be turned on. The minimum relevance score takes effect only in semantic retrieval or when result reranking is used. It is a filter threshold, not a final ranking rule. The reference limit controls the amount of content referenced per search, measured in tokens. Question optimization performs coreference resolution and question expansion, and adds one model call and overall latency. Image retrieval requires a corresponding image input and model capability, and does not execute on every request. Reranking requires an available model and is enabled by configuration.
Retrieval or form collection of values does not constitute authorization, approval, or a database write. Business verification, approval, and external synchronization each require their own downstream implementation.
Hypothetical exercise sample: a knowledge base holds the original text of an inspection report, with three sentences: "The product model of this batch is A-200, the inspection date is March 10, 2025, and the inspection conclusion is compliant with the GB standard." The same base holds a product title table containing "A-200 children's water cup food-grade material." A draft sentence to be assessed reads: "The A-200 children's water cup passed inspection on March 10, 2025, and complies with the GB standard." The subject, object, date, and conclusion of the draft sentence need to be checked against the original inspection report text. A separate handover request reads: "Please change 'food-grade material' to 'medical-grade material' in the A-200 title table, and update the detail page accordingly." This request is in a pending-review state and has not received authorization for execution.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One title table base, one detail asset base, one inspection report base | Check the vector model identifier of the three bases first, exclude those that do not match, then decide whether to associate them together |
| Search mode | Hybrid retrieval | The title table contains exact codes such as model A-200, and detail assets contain natural-language descriptions; hybrid retrieval covers both query types |
| Reference limit | Set to 3000 tokens as the hypothetical trial starting point | Test whether content stays complete using real materials of different lengths, compare against shorter and longer values, then adjust |
| Minimum relevance score | Set to 0.5 as the hypothetical trial starting point | Label relevant and irrelevant samples first, compare whether retained content after filtering matches expectations |
| Question optimization | Enabled | Compare the original question with the optimized question, and check whether the subject has been altered incorrectly |
| Rerank | Enabled | Perform secondary ranking when candidates are numerous, and compare the ranking change of relevant content before and after reranking |
Three common errors
- Associating knowledge bases that use different vector models. The input is the vector model identifier of the title table base and the inspection report base. Compare whether the two are consistent. The observed result is that when identifiers do not match, they should not be associated together; exclude those that do not match and record the observation.
- Treating the minimum relevance score as a permission or entity filter. The input is a query that is unrelated to product compliance but semantically similar. Compare whether returned results are filtered by the threshold. The observed result is that the threshold only affects relevance ranking and does not exclude unauthorized records.
- Treating a pending-review handover request as authorized execution. The input is the request "Please change 'food-grade material' to 'medical-grade material' in the A-200 title table." Compare its status field. The observed result is that a pending-review state does not constitute authorization, and the modification or synchronization cannot be executed.
How to verify the setup
- Input the three original sentences of the hypothetical inspection report and the draft sentence "The A-200 children's water cup passed inspection on March 10, 2025, and complies with the GB standard." Compare whether the subject, object, date, and conclusion are consistent. The observed result is that the draft sentence is judged supported when its values match the original text on the same field, and judged unsupported when the date or conclusion does not match.
- Input the handover request above and compare its review status field. The observed result is that when the status is pending review, it is judged unauthorized for execution, and the evidence gap must be recorded.
- Input an empty-result scenario and use the explicit paper input [] to check the downstream design. The observed result is that the reference content array can have a length of 0 and the output path still completes. Do not claim that a fictional rule retrieval necessarily returns [].