What the documents look like
JD POP third-party store operations revolve around open-platform dispute rules and warranty-period service documents. Working materials come from the merchant backend, the platform help center, open-platform rules, and logistics tracking. Day-to-day documents include product information tables, dispatch-time commitments, dispute handling forms, and warranty-period service records. These materials update daily or per order.
For operations handover, the relevant content splits into four groups: product information, time commitments, dispute forms, and service records. Permission status, operation logs, and pending requests are scattered across these documents. A retrieval result presents cited content only. It does not show that a permission was granted or that an action was carried out. Dispatch-time commitments and dispute handling forms often contain date, address, and status fields, and retrieval needs to preserve those field values as written.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together in one node. The model identifier must be checked first, and any base that does not match is excluded.
The node outputs cited content as an array, and the length can be 0. When no content is found, the downstream path still runs with an empty reference. Search mode can be semantic search, full-text search, or hybrid search. Hybrid search merges multiple retrieval paths through RRF. RRF is a fusion step inside hybrid search, not a switch that is turned on.
Minimum relevance applies only under semantic search or when result reranking is used. It is a filter threshold. It does not perform permission separation or entity separation. The reference cap is measured in tokens and limits how much content a single search may cite. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. Image retrieval requires image input and matching model capability. Reranking requires a rerank model and must be enabled by configuration.
Retrieval is an information-fetching step that sits outside authorization. It does not decide who holds a permission.
Configuration and implementation checks
The table below concerns one hypothetical drill sample for this page. It lists only documented settings or implementation checks tied to this feature and the handover task.
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the base that holds the product information table and the dispatch-time commitment and uses the same vector model | Check the vector model identifier first and exclude bases with a different model; compare whether retrieval output still comes from that base before and after linking |
| Search mode | Hybrid search | The sample contains natural-language questions and numbered fields; compare cited content from semantic, full-text, and hybrid modes on the same question |
| Reference cap | 800 tokens | A hypothetical starting point because sample materials vary in length; compare content completeness against 400 and 1200 tokens |
| Minimum relevance | 0.35 | A hypothetical starting point that applies only under semantic search or reranking; compare filtering of relevant and irrelevant samples against 0.2 and 0.5 |
| Question optimization | Enabled | The sample contains an unclear follow-up question; compare whether the optimized question keeps the correct subject |
| Result reranking | Enabled | The sample has many candidates and a clear question; compare the 0-1 score ordering before and after reranking |
Three common errors
- Treating a retrieval hit as permission confirmation. Input a dispute handling form whose status is "pending review," then compare the cited content against the authorization field. If the citation shows only the form status and no approval conclusion, the permission state remains unsupported, and the record cannot be read as authorized.
- Linking knowledge bases across different vector models. Input two bases with different vector model identifiers, compare the identifiers, and exclude any that do not match. Do not run a cross-model retrieval to compare missed recalls; record what is observed.
- Treating minimum relevance as an entity filter. Input a query about an unrelated business matter and compare cited content before and after filtering. If unrelated records are still returned, the threshold does not carry permission or entity separation, and business filtering needs a separate downstream implementation.
How to verify the setup
The three checks below use a hypothetical drill sample. Each states the input, the comparison, and the observable result. No test has been run.
Hypothetical drill sample: the source text reads, "Product information table states: Product A, warranty period 12 months, dispatch-time commitment is dispatch within 48 hours after payment. Dispute handling form states: Order B, buyer address in a certain province and city, status pending review." The draft sentence under review reads, "The dispute for Order B has been approved and a refund may be executed." That draft sentence does not match the status field in the source text.
- Verify whether retrieval preserves the status field. Input the question "dispute handling status for Order B" and compare whether the cited content contains the words "pending review." If the citation contains that phrase and does not contain "approved," retrieval has preserved the original status, and the draft sentence's "has been approved" is unsupported.
- Verify whether the reference cap affects content completeness. Input the question "warranty period and dispatch time for Product A" and compare whether both the warranty period and the dispatch time are returned at a reference cap of 800 tokens. If both appear, the content is complete. If only one appears, adjust the cap and compare again.
- Verify the empty-result path. Input the question "nonexistent order number C" and compare whether the output is an array of length 0. If the output is 0 and the downstream design can accept an empty reference, the empty-result path runs through. Do not claim that a query for a fictional rule must return an empty array.