What the documents look like
JD POP third-party store operations work with four recurring document types: product information sheets, dispatch timeliness agreements, dispute handling forms, and warranty-period service records. These materials come from the JD merchant backend, the JD help center, JD open platform rules, and logistics tracking. Updates occur daily or per order.
The comparison task places agreed text and actual records for the same order or the same after-sales matter side by side, then checks whether the subject, object, status, and valid scope match. Materials sit in several sources, so the retrieval step first recalls candidate passages and hands them to a person for field-by-field comparison. Retrieval itself does not decide who is responsible, whether a dispute is resolved, or whether an authorization is valid.
A hypothetical exercise sample is used throughout this page. A POP store operator is checking after-sales responsibility for one order. The knowledge base holds three source sentences. Sentence one: "The product information sheet states the warranty period for this product is twelve months, counted from the date of receipt." Sentence two: "The dispatch timeliness agreement states the seller should dispatch within forty-eight hours after buyer payment." Sentence three: "The dispute handling form states the buyer raised an after-sales request on the tenth day after receipt, and the handling status is pending review." The draft sentence under review reads: "After-sales responsibility for this order has been borne by the seller and was completed within the warranty period." None of the three source sentences states a responsibility conclusion or a completion fact. The phrases "has been borne by the seller" and "was completed" have no corresponding field in the source text.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection only classifies data and does not affect search results. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges the two paths with RRF. The minimum relevance value takes effect only under semantic retrieval or when result re-ranking is used. The citation limit is measured in tokens. Retrieval recalls candidate content; it does not perform permission checks and does not replace a person confirming authorization status.
An empty reference output is possible. The reference content is returned as an array whose length can be zero, and the output path still completes when nothing is found. This means a downstream design must handle a zero-length array rather than assume content always arrives.
Re-ranking requires a re-rank model and must be enabled according to settings. Image retrieval requires an image input and matching model capability. These optional paths are not executed on every request. Knowledge bases using different vector models cannot be associated together; the model identifier is checked first and non-matching bases are excluded. No runtime outcome is asserted for an attempted cross-model association.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the single knowledge base holding the four material types, all on one vector model | Cross-model bases cannot be associated; check the vector model identifier first and exclude non-matching bases |
| Search mode | Start with hybrid retrieval | Materials contain periods and numbers such as twelve months and forty-eight hours, so semantic and keyword paths both matter |
| Citation limit | Set to a token amount that can hold the three source sentences | Test with actual material lengths to see whether content stays complete, then adjust up or down |
| Minimum relevance | Start at a middle value, label relevant and irrelevant samples, then adjust | This value takes effect only under semantic retrieval or result re-ranking; compare filtering against labeled samples |
| Question optimization | Enable first, then compare the original question with the optimized question | Used for coreference resolution and question expansion; check whether the subject was altered |
| Result re-ranking | Leave disabled first | Re-ranking needs a model and must be enabled by settings; with few candidates, observe the merged result first |
Three common errors
- Treating retrieval recall as authorization confirmation. In the sample, the dispute handling form status is pending review. Retrieval can recall that sentence, but a pending status does not show the seller has borne responsibility. Check: input the source sentence, compare whether the draft sentence contains "has been borne by the seller," and observe whether the two agree.
- Merging different fields into one finding. The twelve-month warranty period and the forty-eight-hour dispatch deadline are separate fields and cannot be derived from each other. Check: input the two source sentences separately, compare whether the draft sentence combines both fields into one conclusion, and observe whether a cross-field inference appears.
- Reading a missing statement as a denial. None of the three source sentences states completion. Check: input the three source sentences, compare whether a field corresponding to "was completed" exists, and observe whether the absence has been written as a settled conclusion.
How to verify the setup
- Input the three hypothetical source sentences and the draft sentence. Compare the responsibility and completion wording in the draft against the source. If no corresponding field exists in the source, the draft is unsupported and the evidence gap is retained.
- Input the pending-review status sentence from the hypothetical sample. Compare whether the draft sentence turns it into a completed state. If the status field still reads pending review, no completion conclusion may be output.
- Input an empty reference array as a paper input. Check whether the downstream design accepts a zero-length reference output. This does not claim that any rule query actually returns an empty array.
The verification record notes actual observations and failure branches. It does not guarantee a first-hit result or full recall. Image retrieval requires an image input and model capability; re-ranking requires a model and must be enabled by settings.