What the documents look like
JD POP third-party store operations run on open-platform dispute rules and warranty-period service records. The working materials are product information sheets, dispatch-time commitments, dispute handling forms, and warranty service records. In cross-border materials and multilingual retrieval, one item can exist as a Chinese original and a foreign-language translation at the same time. Fields in the product information sheet, clauses in the dispatch-time commitment, and evidence entries in the dispute handling form each need their source checked line by line. The retrieval step matches a user question against knowledge base content by vector or full-text search and returns a citation array whose length may be zero. That citation array only shows which content was matched. It does not show that a material has passed business review, received authorization, or taken effect.
Two industry facts shape this page. First, a single material may be present in more than one language, so tracing a citation means returning to the specific original text rather than to a paraphrase. Second, retrieval output and business status are separate facts: a dispute handling form can be retrieved while its processing status remains pending review.
Constraints at the citation tracing step
Tracing requires that every citation in an answer can be returned to a specific material original, and that the material version and validity scope are checked. The knowledge base search node can associate one or more knowledge bases that share the same vector model. Knowledge bases with different vector models cannot be associated together. During checking, confirm the knowledge base identifier and the vector model identifier first, exclude those that do not match, and do not run cross-model retrieval to compare missed recalls. Citation content is output as an array, and the output path runs through even when no content is found, so an empty array does not mean the workflow errored out and stopped. The minimum relevance value in semantic retrieval is a filter threshold; it does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external business registry sites or other systems. Image retrieval needs a corresponding image input and model capability, and reranking needs a model and must be enabled by configuration. The optional paths drawn in the document flow diagram do not mean every request executes all of them.
Hypothetical exercise sample. Material original: "This store's product information sheet (version V3) states the dispatch time commitment as dispatch within 48 hours after payment, applicable from 1 March 2025 to 31 March 2025." "Dispute handling form number JF-2025-0312 records that the buyer claimed dispatch did not occur within the agreed time; the seller submitted a logistics tracking screenshot; processing status is pending review." The draft sentence to be judged: "The store completed dispatch under the dispatch time commitment, so the dispute does not hold." Checking compares the draft sentence against the material original. The material states only the dispatch time commitment and the pending review status. It does not state the actual dispatch time, and it does not state a dispute conclusion. The draft sentence exceeds the original's scope and is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Two knowledge bases with the same vector model | Check knowledge base identifiers and vector model identifiers first, and exclude those with a different vector model |
| Search mode | Hybrid retrieval | Compare against semantic retrieval and full-text retrieval separately on identifier-type and natural-language questions |
| Citation limit | 3000 tokens as the hypothetical starting value | Compare against the unadjusted value and observe whether cited content length covers the material original |
| Minimum relevance | 0.3 as the hypothetical starting value | Compare against 0.5 and observe whether low-relevance fragments are filtered |
| Result reranking | Enabled | Compare against disabled and observe whether ordering changes when candidates are numerous |
| Response format | Keep default | Compare against json_schema and confirm whether a downstream step needs a fixed structure |
Three common errors
- Treating a retrieval match as business authorization. Input is a dispute handling form with pending review status. Compare the retrieval output against the processing status field and observe whether pending review is written as upheld or rejected. Pending review does not show that an operating permission is absent, and it does not show a dispute conclusion.
- Treating an empty citation array as an error. Input is a paper empty array []. Compare the downstream design and observe whether it is designed to continue. An empty array means no content was found; it does not mean the workflow stopped.
- Associating knowledge bases with different vector models. Input is two knowledge base identifiers. Compare their vector model identifiers and observe whether non-matching ones are excluded. Cross-model association is not allowed. Do not then infer missed recalls or redundant returns as runtime outcomes.
How to verify the setup
- Input the material original and the draft sentence from the hypothetical sample. Compare whether the draft sentence exceeds the subject, object, status, and validity scope stated in the original, and observe whether the draft sentence is marked unsupported. The expectation is that it is marked unsupported. This does not claim that a test was run.
- Input the paper empty array []. Compare whether the downstream node is designed to accept an empty citation and observe whether the output path runs through. The expectation is that it continues. This does not claim that a constructed rule retrieval necessarily returns [].
- Input two knowledge base identifiers. Compare their vector model identifiers, observe whether those with different vector models are excluded, and record the actual observation. The expectation is that non-matching ones are excluded. Do not run cross-model retrieval to compare missed recalls.
Fields not stated in the documents remain unknown, and missing evidence is recorded as a gap.