What the documents look like
Third-party overseas warehouses can serve several sales channels at once, and the seller allocates stock across them. This differs from a single-channel replenishment model, where the platform manages the inventory pool. Records for this category come from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems, and carrier bills. The working documents include replenishment plans, inbound appointment requests, inventory age reports, and last-mile delivery orders, refreshed weekly or daily.
For retrieval, the practical difficulty is that one document mixes languages. A single replenishment plan may carry an English SKU, a Chinese remark, and a destination port code. A question phrased in natural language may need to reach a passage written in another language, while a question built around an exact SKU or model number depends on keyword matching. The knowledge base search node can associate one or more knowledge bases that use the same vector model, and the referenced content returns as an array whose length can be 0, so the downstream path still runs when nothing is found.
Constraints at the retrieval step
Search mode is compared against the test question rather than assumed. Semantic retrieval matches by vector distance and can cross languages, but its precision depends on model training and on how complete the indexed text is. Full-text retrieval depends on keyword hits and suits part numbers, model codes, and proper nouns. Hybrid retrieval uses both and merges the two result sets through RRF, which the documentation describes as a fusion step, not a separate switch. Where candidates are numerous, a ReRank model reorders them and produces a 0-1 score that can drive filtering.
The citation cap is measured in tokens, not in a top-k item count, because chunk lengths vary widely in mixed knowledge bases. The minimum relevance value is a 0-1 filter threshold and applies only under semantic retrieval or when result reordering is in use. It is a threshold, not a final sort rule, and it does not separate unrelated businesses or unauthorized records; retrieval is not authorization. Question optimization performs coreference resolution and question expansion before the search, which adds a model call and helps multi-turn follow-ups, while exact identifier questions may not benefit. Reordering requires an available model and must be enabled by configuration. Image retrieval requires an actual image input and matching model capability, and the optional paths drawn in the overall flow diagram do not all execute on every request.
Hypothetical paper exercise sample: an overseas warehouse operator holds a replenishment plan whose source text reads, "On 10 March 2025, SKU A-102 is planned for replenishment to the German warehouse, quantity 600 units, inbound appointment status pending review, last-mile delivery order not yet generated." A draft sentence under review reads, "This replenishment plan has been approved and inbound scheduling may proceed." The question is whether the draft sentence agrees with the source text. The source states only that the appointment status is pending review and does not record any approval fact.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one replenishment knowledge base using the same vector model | Check the vector model identifier of each base, exclude those that differ, then run the search |
| Search mode | Test a multilingual question under semantic retrieval first, then test an SKU number under full-text retrieval | Compare the retrieved content for the same question across the two modes |
| Citation cap | Set a token amount that covers the full set of fields in one replenishment plan | Test plans of differing length and check whether fields are truncated |
| Minimum relevance | Begin at a middle value, label relevant and irrelevant samples, then adjust up or down | Applies only under semantic retrieval or reordering; observe whether unrelated fragments are filtered |
| Question optimization | Enable for follow-up questions containing references, disable for explicit identifier questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reordering | Enable when candidates are numerous, filter by the reorder score | Confirm the reorder model is available, compare ordering before and after |
Three common errors
- Reading "pending review" as "approved." In the sample, the source text states the inbound appointment status is pending review, while the draft sentence states approval has been granted. Both address the same subject and the same field with mutually exclusive values, so the two conflict. The error is treating a pending state as an approval.
- Treating the similarity score as a permission filter. The minimum relevance value is a filter threshold and not a final sort rule, and it does not exclude unrelated business lines or unauthorized records. The error is assuming that raising relevance isolates replenishment records belonging to other channels.
- Associating knowledge bases that use different vector models. The documentation requires associated knowledge bases to share the same vector model, so cross-model association is not permitted. The error is selecting bases with different models and expecting retrieval to proceed normally.
How to verify the setup
- Input the sample source text and the draft sentence, then compare whether "pending review" and "approved" occupy the same field with mutually exclusive values. The expected observation is that the two conflict, the draft sentence is not supported, and the evidence gap remains open.
- Input a retrieval question containing an SKU number and run it under semantic retrieval and under full-text retrieval. Compare whether the retrieved content includes the record for that number. The expected observation is that full-text retrieval hits the number, while the semantic result must be checked against the same record.
- Input a paper retrieval question with no corresponding material and check whether the referenced content output is an array of length 0. The expected observation is that the output length can be 0 and the downstream path still runs. Missing evidence stays missing; pending review does not demonstrate a lack of operating authority, and retrieval does not constitute authorized execution.