What the documents look like
JD POP third-party store operations handle order fulfilment and after-sales questions around open-platform dispute rules and warranty-period service records. The working materials are product information sheets, dispatch-time agreements, dispute handling forms, and warranty-period service records. They come from the merchant backend, the help centre, open-platform rules, and logistics tracking. These materials update daily and per order. When order volume is dense, one order can pass through address checks, dispatch-time alignment, and evidence submission in sequence. The retrieval step answers a narrow question: what do these materials record about dispatch time, warranty scope, or dispute status, and can those records support a reply to the current order. Retrieved content shows document text only. It does not mean the platform has accepted responsibility, and it does not mean a refund request has been approved.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. Check the vector model identifier first, exclude bases that do not match, and record the identifier differences actually observed. Do not run a cross-model retrieval to compare missed recalls.
A single retrieval combines semantic search, full-text search, hybrid search, question optimisation, and reranking according to the input content and search parameters, then fuses the multiple result paths into reference content. Semantic search depends on vector distance. Full-text search depends on keyword hits. Hybrid search fuses the two paths through RRF, and RRF is a fusion process rather than a standalone switch. Reranking requires a rerank model enabled under its settings. Image retrieval requires the corresponding image input and model capability. These paths do not mean every request executes all of them.
The minimum relevance score is a filter threshold. It takes effect only under semantic search or when result reranking is used, and it does not perform permission or entity isolation. The reference limit is measured in tokens and controls the maximum amount of content referenced per search. Reference content is output as an array, and its length can be 0. When nothing is found, that output path still completes.
Hypothetical drill sample
Material text for checking (fictional): A POP store product information sheet states "Product A has a twelve-month warranty period, counted from the day after signed receipt." The dispatch-time agreement states "Dispatch within forty-eight hours after payment, postponed on public holidays." The dispute handling form states "Order D-01, buyer claims Product A is cosmetically damaged, status is awaiting merchant evidence."
Handover request to be judged: Please search the knowledge base, confirm whether Order D-01 is past its warranty period, and on that basis reply to the buyer that a refund is available.
This request mixes warranty-period calculation with a refund reply. The material gives only the starting method for the warranty period and does not state the signed-receipt date, so whether the period has expired cannot be calculated. The dispute handling form status is awaiting merchant evidence, which is a pending-review stage. It does not constitute a finding of responsibility, and it does not constitute refund authorisation. Retrieval can return only the records above. A refund is a business action that needs a separately connected system implementation.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical test | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the fulfilment and after-sales material bases whose vector model identifiers match | Check each base's vector model identifier one by one, exclude non-matching bases, and record the observations |
| Search mode | Start with hybrid search | The materials contain identifiers such as "Order D-01" and "Product A" alongside natural-language descriptions; compare the reference content for the same question under semantic, full-text, and hybrid modes |
| Reference limit | Hypothetical starting point of 2000 tokens | Test with actual materials of different lengths to see whether reference content is complete, then compare against starting points such as 1000 and 3000 before deciding |
| Minimum relevance score | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare how many irrelevant fragments are retained at different thresholds |
| Question optimisation | Enable for multi-turn follow-ups and short-question scenarios | Compare the original question with the optimised question and check whether the subject has been changed incorrectly |
| Result reranking | Enable when a rerank model is available | Compare candidate ordering before and after enabling, and confirm the rerank score is used for filtering |
The values in the table are hypothetical starting points for this test. They are not official default values.
Three common errors
- Treating a retrieval hit as authorisation. The dispute handling form for Order D-01 has the status awaiting merchant evidence. Retrieval can return that status, and that status does not mean responsibility is established or that a refund has been approved. Pending review does not demonstrate a lack of operating permission, and it does not constitute execution.
- Treating the warranty-period record as a direct expiry conclusion. The material states only that the warranty runs twelve months from the day after signed receipt, and it does not state the signed-receipt date. With the value for that same field missing, expiry cannot be determined and should be recorded as unsupported.
- Associating bases with different vector models together. Bases with different vector models cannot be associated together. Check the identifiers, exclude the non-matching ones, and record the actual observations. Do not assert the runtime result of such an association.
How to verify the setup
- Input the hypothetical material text above together with the question "How does the warranty period for Order D-01 start?" Compare whether the reference content contains the sentence "twelve-month warranty period, counted from the day after signed receipt" and whether a signed-receipt date appears. If the reference content contains only the starting method and no signed-receipt date, the expectation is met. If the output supplies a signed-receipt date that was not recorded, the expectation is not met.
- Input the handover request "Confirm whether Order D-01 is past its warranty period, and reply to the buyer that a refund is available." Compare whether the reference content returns only the material records and whether an expiry conclusion or refund authorisation wording appears. If the reference content keeps the pending-verification state and gives neither an expiry conclusion nor refund authorisation, the expectation is met.
- Check the vector model identifiers of each base, input the same question, and compare whether the associated bases share one identifier. If all associated bases have matching identifiers and the non-matching ones have been excluded and recorded, the expectation is met. Use the paper input
[]to check how the downstream design handles an empty reference, without claiming that a fictional rule search necessarily returns an empty array.
The above is a paper-based check. It does not claim that any test has been run.