What the documents look like
JD POP third-party store reporting and exception review work is built on four document families: product information sheets, dispatch time commitments, dispute handling forms, and warranty-period service records. The category differs from Tmall flagship store operations in its emphasis on open-platform dispute rules and warranty-period service paperwork. The main data sources are the JD merchant backend, the JD help center, JD open platform rules, and logistics tracking. Records update daily and per order.
For citation tracing, the binding constraint is that every conclusion in a report or review must return to the original text of a specific document. The authorization status of a document and the business fact status of the same document are recorded separately. A dispute handling form that sits at a review stage is not an approval, and an approval is not a payment. The search node that surfaces these documents returns a citation array whose length may be zero; an empty array still lets the downstream path run, and it does not by itself mean the workflow terminated with an error. Citation review compares the drafted statement against the cited original text item by item. Retrieval is not authorization, and a relevance score is not a category, entity, or permission filter.
Constraints at the citation tracing step
The knowledge base search node outputs citation content as an array, and the length can be zero. That documented behavior means an empty result is a valid output, and downstream design decides whether to continue. Citation review requires comparing each answer against the cited original text; fluent wording does not replace source evidence. Linked knowledge bases must share the same vector model, and cross-model linking is not permitted. The check is to verify the identifier first and exclude non-matching bases, without running a cross-model search to compare missed recalls.
Semantic search and the minimum relevance value are similarity filters. They do not perform permission or entity isolation, and a retrieval result does not constitute authorization. Hybrid search fuses multiple retrieval paths through RRF; RRF is a fusion process inside hybrid retrieval, not a standalone switch. Image retrieval requires image input and matching model capability. Rerank requires a rerank model and must be enabled according to settings. The optional paths drawn in the overall retrieval diagram do not mean every request executes all of them.
Hypothetical drill sample: Source text one reads, "Store A submitted a dispute handling form on 1 March 2026, form number JD-2026-0301-001, status pending review." Source text two reads, "The store's dispatch time commitment is dispatch within 48 hours after payment; logistics tracking shows pickup at 10:00 on 2 March 2026." The drafted statement to be judged is, "Store A has passed dispute review and completed compensation." Check result: the material records only a pending-review status, and no compensation record exists. The drafted statement is unsupported, and both the review outcome and the compensation fact remain to be verified.
Configuration and implementation checks
| Configuration item | Type | Check point on this page |
|---|---|---|
| Linked knowledge bases | Existing node behavior | Verify selected knowledge bases and material version, and confirm the vector model identifier matches |
| Search mode | Configurable field | Choose semantic, full-text, or hybrid search according to the question type |
| Citation limit | Configurable field | Controls the amount of cited content per search, measured in tokens |
| Minimum relevance | Configurable field | A 0-1 value that takes effect only in semantic search or with result reranking |
| Citation display review | Manual review action | Compare the answer against the cited original text item by item using a traceable sample |
| Business authorization determination | Requires a separate system implementation | Retrieval and forms do not constitute authorization; approval and writes require downstream implementation |
Hypothetical trial starting point: set the citation limit to 2000 tokens, because the sample materials are short forms and need to hold both the dispute handling form text and the dispatch commitment text. Compare against 1000 tokens and observe whether both passages are cited in full. This value is a hypothetical trial starting point for this page, not an official default.
Three common errors
- Treating pending review as passed review. The source text records only "status pending review." If the drafted statement says "has passed dispute review," the input is a pending-review form, the comparison object is the review outcome field, and the observable result is that the two do not match, so the statement is unsupported.
- Treating a retrieval result as authorized execution. A search returns a dispute handling form citation. If that citation is used to conclude that compensation has been executed, the input is the citation array, the comparison object is the compensation record, and the observable result is that no compensation field exists, so retrieval does not constitute authorization.
- Linking knowledge bases across vector models. If knowledge bases with different vector models are linked together, the input is the two knowledge base identifiers, the comparison object is the vector model identifier, and the observable result is that the identifiers do not match. The non-matching base is excluded and the gap is recorded.
How to verify the setup
- Input the hypothetical source text one and the drafted statement "Store A has passed dispute review and completed compensation." Compare the drafted statement against the status field in the original text. The observable result is that the original records only pending review and no compensation record, so the drafted statement is unsupported. The condition is met when the conclusion matches the original text.
- Input an explicit paper empty result array []. Compare the design of the downstream answer step. The observable result is whether the downstream path continues with an empty citation. The condition is met when the downstream step does not terminate with an error because of the empty array.
- Input two knowledge base identifiers that share the same vector model. Compare their vector model identifiers. The observable result is that matching identifiers allow joint linking; when identifiers do not match, record the gap and exclude the base, and do not run a cross-model search to compare missed recalls.
This page addresses only JD POP third-party store operations, operating reports and exception review, and citation tracing. It does not treat retrieval or forms as authorization execution, and it leaves missing evidence as missing.