What the documents look like
Title and detail compliance work handles advertising truthfulness and qualification display. It does not handle parcel delivery. The materials behind this work come from platform rule centers, advertising-law compliance checklists, brand asset libraries, and inspection-agency reports. Day-to-day documents include product title tables, detail page assets, advertising prohibited-word lists, and quality inspection reports. The most time-consuming steps are checking absolute-language claims, comparing physical parameters, and replacing non-compliant assets. These materials shape what a retrieval library must cover: text semantics carry most of the load, while item numbers, model designations, and inspection-item names need exact keyword hits.
Two features of this material set affect retrieval directly. First, the same product can appear in a title table, a detail page, and an inspection report, and each document carries different fields. Second, prohibited-word rules govern titles, while inspection reports govern measured parameters. A retrieval hit on one document type cannot stand in for evidence about another.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Vector search uses top-k recall. Semantic retrieval depends on vector distance; full-text retrieval depends on keyword hits; hybrid retrieval merges both paths through RRF. The citation limit is measured in tokens. The minimum relevance threshold takes effect only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and matching model capability. Reranking requires a model and must be enabled by configuration. These are optional paths; they do not all run on every request. Retrieval results do not carry permission or entity-isolation duties.
Hypothetical paper exercise. Source text one: a rice cooker detail page states "rated capacity 3 L, rated power 500 W." Source text two: the advertising prohibited-word list states "absolute terms must not be used in product titles." Source text three: an inspection report states "sample name: rice cooker, inspection item: input power, conclusion: pass." Draft sentence to be assessed: "The rice cooker title uses the phrase 'most power-saving on the whole network'; the detail page parameters match the inspection report, so it can be submitted for listing review." Assessment points: the draft sentence concerns one product, and its fields are title wording and detail parameters. The prohibited-word list governs titles; the report governs input power. These are different fields, and neither can support or refute the other. Whether listing review is approved remains unverified; a retrieval hit does not constitute authorization.
Configuration and implementation checks
| Configuration item | Value or check on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Select title and detail compliance libraries that use the same vector model | The node requires the same vector model; check the model identifier first and exclude mismatches |
| Search mode | Compare semantic, full-text, and hybrid per test question | Item and model questions use full-text or hybrid; natural-language questions use semantic or hybrid |
| Citation limit | Hypothetical trial starting value of 3000 tokens | This value is only a paper comparison starting point for this exercise; compare against 1500 and 6000 for content completeness before adjusting |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold | Takes effect only under semantic retrieval or result reranking |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject was altered |
| Reranking | Use when a model is available and the setting is enabled | Rerank scores range from 0 to 1 and can be used for filtering |
Three common errors
- Treating a retrieval hit as listing authorization. Input the draft sentence; the system returns a prohibited-word list entry. The comparison action is to check whether the returned content contains an authorization conclusion. The observable result is that only rule text is returned, with no approval status, so listing review remains unverified.
- Forcing an association across vector models. Input two knowledge base identifiers that use different vector models. The comparison action is to check whether the model identifiers match. The observable result is that a mismatch leads to exclusion; no cross-model retrieval is run to compare missed recalls.
- Using different fields to refute each other. Input the detail page power parameter and the inspection report input-power conclusion. The comparison action is to check whether subject, period, and field are the same. The observable result is that the fields differ, so the record can only be noted as unsupported, not as a contradiction.
How to verify the setup
- Input source text one and the draft sentence. Compare whether the detail page parameter and the parameter stated in the draft sentence share the same subject and field. The observable result is that when the parameters match, that part is supported; the title wording portion still returns to the prohibited-word list for separate checking.
- Input source text two. Compare whether the title wording falls within the prohibited-word list scope. The observable result is that when an absolute term is hit, that expression is unsupported; when there is no hit, the record stays unverified, and no listing permission is inferred.
- Input an empty citation array. Compare whether downstream nodes continue executing along the citation path with length 0. The observable result is that the path running through matches the documented behavior; do not claim that a fabricated rule query necessarily returns an empty array.