What the documents look like
E-commerce warehouse receiving and putaway handles the first physical-and-record confirmation when goods enter the warehouse. The working materials are the purchase arrival note, the receiving inspection sheet, the putaway task, and the discrepancy report. Their sources are the WMS, ERP purchasing records, supplier delivery notes, and barcode scans. Staff on the receiving dock and the WMS administrator consult these materials during store service and internal training, so retrieval has to answer questions about acceptance standards, discrepancy handling, and putaway rules, and it has to land on a specific batch and document number rather than a general policy paragraph.
These materials change per batch and daily. That frequency matters here only in one direction: a retrieval result that cannot be tied to a batch and a document number is not useful for the training question being asked, because the underlying facts differ from one arrival to the next.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together. The practical step is to check the vector model identifier of each candidate library and exclude the ones whose identifiers do not match; a cross-model retrieval is not run in order to compare what it misses.
Retrieval runs as semantic search, full-text search, or hybrid search. Hybrid search combines the multiple retrieval paths through RRF, and RRF is the fusion process applied to those results, not a switch that is turned on. The minimum relevance value applies only under semantic search or when result reranking is used. That value is a filtering threshold; it does not perform permission or entity isolation. The reference cap is measured in tokens. Question optimization performs coreference resolution and question expansion and adds one model call before retrieval. Image retrieval requires an image input and matching model capability. Reranking requires a rerank model and has to be enabled according to the setting. The optional paths shown in the overall retrieval diagram do not mean every request executes all of them. An empty result still returns through the output link, and the reference content array length can be zero.
Hypothetical drill sample: the source text to be checked reads, "Purchase arrival note PO-2026-0417 states an arrival quantity of 120 pieces; receiving inspection sheet YS-2026-0417 records 118 pieces actually received; discrepancy report CY-2026-0417 notes a shortage of 2 pieces, status pending supplier confirmation." One draft sentence awaiting judgment reads, "The supplier has confirmed this shortage and the inventory adjustment has been completed." Compared with the source text, the source states only that supplier confirmation is pending; it does not state that confirmation was completed or that an inventory adjustment was executed.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only receiving and putaway libraries that use the same vector model | Check each library's vector model identifier and exclude libraries whose identifiers differ |
| Search mode | Run semantic search and full-text search separately | Put the same question to both modes and compare the reference content returned |
| Reference cap | Hypothetical starting point of 800 tokens | Test with inspection sheets and discrepancy reports of different lengths for content completeness, and compare against a smaller value |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then observe which items survive filtering |
| Question optimization | Enable for follow-up questions that contain references | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enable according to the setting when candidates are numerous | Where a rerank model is available, compare the rerank score with the vector score |
Three common errors
- Treating a retrieval result as an authorization basis. Retrieving discrepancy report CY-2026-0417, which notes pending supplier confirmation, does not mean authorization exists to execute an inventory adjustment. A pending-review state does not demonstrate that operating permission is absent, and it does not demonstrate that approval has been granted.
- Associating knowledge bases that use different vector models. The input is two library identifiers with different vector model identifiers; after comparing the identifiers, the non-matching library is excluded. No further inference is drawn about how a joint association would behave.
- Using the reference cap interchangeably with a thinking budget. The reference cap is measured in tokens and controls how much content each search cites. A thinking budget is a separate kind of setting and is not used as a reference cap.
How to verify the setup
- Enter the hypothetical source text "Purchase arrival note PO-2026-0417 states an arrival quantity of 120 pieces; receiving inspection sheet YS-2026-0417 records 118 pieces actually received; discrepancy report CY-2026-0417 notes a shortage of 2 pieces, status pending supplier confirmation," and compare it with the draft sentence "The supplier has confirmed this shortage and the inventory adjustment has been completed." If the reference content returned by retrieval supports only a pending-confirmation status, the draft sentence is not supported and the evidence gap is recorded.
- Enter the identifiers of two libraries that use different vector models, and compare whether the vector model identifiers match. If the identifiers do not match, exclude that library and do not run a joint association retrieval; record what was actually observed.
- Enter an explicit paper input of an empty result, [], and check whether the downstream link continues to execute. If the reference content length is zero and the output link runs through, the empty-result handling matches the documented description. No claim is made that a fictional rule retrieval necessarily returns [].