What the documents look like
JD POP third-party store operations work with open-platform dispute rules and warranty-period service records. The day-to-day materials are product information tables, dispatch-time agreements, dispute handling forms, and warranty-period service records. In cross-border and multilingual retrieval, the same materials commonly exist in Chinese alongside target-market languages. Retrieval demand concentrates on two question types: identifiers, model numbers, and proper nouns, and natural-language phrasing. Materials update daily and per order. This page addresses knowledge-base retrieval and recall for a single query; it does not depend on batch synchronization mechanisms.
Constraints at the retrieval step
The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. During implementation, check the vector model identifier first, exclude those that do not match, and record what is actually observed. Do not run cross-model retrieval to compare recall differences.
Semantic retrieval relies on vector distance and suits natural-language questions, with documented cross-multilingual understanding. Full-text retrieval suits precise queries such as identifiers, model numbers, and proper nouns. Hybrid retrieval uses both recall paths and merges results through RRF. RRF is a fusion process, not a switch. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold. The citation limit is measured in tokens. Question optimization handles coreference resolution and question expansion, and adds one model call. The retrieval output is an array whose length can be zero; the empty-result path still completes.
Hypothetical paper exercise sample: Source text one reads "Order A-1024, dispatch-time agreement requires dispatch within 48 hours of payment, warranty-period service record shows a signed receipt date of 1 March 2024." Source text two reads "Dispute handling form D-77, buyer claims dispatch did not occur within the agreed time, seller submitted a logistics tracking screenshot on 5 March 2024." The draft sentence to be judged is "Order A-1024 was dispatched within the agreed time." When checking, compare the dispatch-time agreement text, the logistics tracking time, and the draft sentence for consistency of subject, period, and field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the cross-border materials base that uses the same vector model | Check the vector model identifier and exclude bases on different models |
| Search mode | Test natural-language questions with semantic retrieval first | Compare full-text retrieval results using identifier-type questions |
| Citation limit | Set to a token amount covering one complete dispatch-time agreement | Compare content completeness using materials of different lengths |
| Minimum relevance | Start at 0.3 for this exercise | Label relevant and irrelevant samples, then adjust up or down and compare |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare subject consistency between the original question and the optimized question |
| Result reranking | Enable when candidates are numerous and compare scores | Check the relationship between rerank scores and question completeness |
These values are hypothetical starting points for this exercise and do not represent official default values.
Three common errors
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a similarity threshold; it does not exclude unrelated businesses or unauthorized records, and retrieval itself does not constitute authorization.
- Forcing knowledge bases on different vector models into one association. The node requires the same vector model; when encountering a different model, check the identifier and exclude it. Do not infer missed recall or redundant results.
- Treating an empty citation output as a failure. The output array length can be zero, and the empty-result path still completes. Check the downstream design against the actual input.
How to verify the setup
- Input hypothetical source text one and the draft sentence, then compare whether the subject, period, and field of the draft sentence match the source. The source states only the dispatch-time agreement and the signed receipt date; it does not state the actual dispatch time. The draft sentence is unsupported, and the evidence gap should be recorded.
- Input an identifier-type question and a natural-language question, then compare the recalled content from semantic retrieval and full-text retrieval. Observe whether the identifier is hit by full-text retrieval and whether the natural-language question is hit by semantic retrieval. First-position hits and full recall are not guaranteed.
- Input an explicit paper-based empty result [] and check whether the downstream path completes. An empty array can be accepted. Do not claim that a fictitious rule query necessarily returns [].