What the documents look like
Third-party overseas warehouses can serve several sales channels, and the seller decides how much stock each warehouse receives. That allocation decision sits behind the records this page works with. Procurement and supplier material for overseas warehouse replenishment typically includes replenishment plans, inbound appointment records, inventory age tables, and last-mile delivery notes. These materials arrive from the overseas warehouse WMS, ERP sales forecasts, freight forwarder systems, and carrier bills, with weekly or daily update frequency. Supplier names, purchase order numbers, SKU codes, inbound appointment numbers, inventory age bands, and delivery note numbers are the main objects that retrieval is expected to locate. Retrieval output feeds judgments about stockouts, slow-moving storage charges, and last-mile delays.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; verify the vector model identifier and exclude non-matching bases rather than running cross-model retrieval to compare missed recalls. Search parameters include the search mode (semantic retrieval, full-text retrieval, hybrid retrieval), the reference cap measured in tokens, the minimum relevance value on a 0-1 scale that applies only in semantic retrieval or when result reranking is used, and question optimization for coreference resolution and question expansion. Hybrid retrieval merges multiple result paths through RRF, and ReRank performs a second ordering of candidate text. Referenced content is output as an array, and its length can be 0, so the empty-result path still completes. Image retrieval requires image input and matching model capability; reranking requires a ReRank model and must be enabled in settings. Retrieval is not authorization, and minimum relevance is a filtering threshold that carries no permission or entity-isolation role.
Hypothetical paper test case: the source material reads "On 10 March 2025, replenishment plan BP-0310 for overseas warehouse A shows SKU-8842 with a suggested replenishment of 200 units, supplier S-001, and purchase order PO-2025-0312 generated but not sent." The draft sentence to be judged reads "PO-2025-0312 has been sent to supplier S-001." Comparing the two, the purchase order was generated but not sent, while the draft sentence states it was sent, which does not match the source status.
Configuration and implementation checks
| Configuration item | Relevant setting or check for this page | Hypothetical trial starting value and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases using the same vector model | Verify the vector model identifier and exclude bases on a different model |
| Search mode | Compare semantic, full-text, and hybrid retrieval against the test question | Use "PO-2025-0312" to test whether full-text retrieval returns a hit |
| Reference cap | Measured in tokens, controls the length of referenced content | Start at 2000 tokens, chosen to be compared against the full length of the replenishment plan |
| Minimum relevance | Applies only in semantic retrieval or when result reranking is used | Start at 0.3, compared against scores of labeled relevant and irrelevant samples |
| Question optimization | Coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a ReRank model and must be enabled in settings | After enabling, compare the rerank score with the vector score |
Three common errors
- Treating a retrieval result as authorized execution. Retrieving that purchase order PO-2025-0312 was generated does not mean it was sent or confirmed by the supplier. Retrieval is information access; business verification, approval, and external synchronization require separate downstream implementation.
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a 0-1 filtering threshold and does not exclude unrelated businesses or unauthorized records. The retrieval library carries no permission isolation function.
- Treating an empty reference output as retrieval failure. Referenced content length can be 0, and the empty-result path still completes. Distinguish "no content retrieved" from "retrieval function abnormal"; the former is a normal output.
How to verify the setup
- Input the source material from the hypothetical sample: "On 10 March 2025, replenishment plan BP-0310 for overseas warehouse A shows SKU-8842 with a suggested replenishment of 200 units, supplier S-001, and purchase order PO-2025-0312 generated but not sent." Compare the draft sentence "PO-2025-0312 has been sent to supplier S-001" against the source status. If the draft sentence claims it was sent while the source says it was not sent, the draft does not match the source.
- Input the search term "PO-2025-0312" and select full-text retrieval mode. Compare whether the returned result contains that purchase order number and its status field. If the returned content contains "generated but not sent," the full-text hit condition is met; if an empty array is returned, record it as a miss and check whether the data has finished training.
- Input a query unrelated to the replenishment plan, such as "signing status of last-mile delivery note PD-0401 dated 1 April 2025." Compare whether the returned result contains that delivery note. If an empty array is returned, record it as no content retrieved; if unrelated content is returned, check the minimum relevance setting and the data chunking.