What the documents look like
Peak-sale warehouse and delivery scheduling is organized around short-term peak capacity and the arrangement of people, vehicles and space. This differs from the steady balancing logic of multi-warehouse inventory synchronization. Historical peak-sale orders, platform campaign registrations, WMS capacity and courier pickup capability form the main data sources. Everyday documents include the peak-sale forecast table, staff shift rosters, courier support plans and daily exception reports, refreshed per campaign and daily.
For store service and internal training, these materials are organized into a knowledge base so frontline staff can look up scheduling conventions, shift rules and courier support conditions. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Search modes include semantic retrieval, full-text retrieval and hybrid retrieval; hybrid retrieval merges multiple result paths through RRF, and reranking uses a ReRank model to reorder candidate text. The citation limit is measured in tokens, and the minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion, and adds one model call.
Constraints at the retrieval step
A search returns cited content as an array whose length may be 0, meaning the output path still runs even when nothing is found. Retrieval reflects what text exists in the knowledge base; it does not carry out the business actions that text describes.
Hypothetical drill sample: before a peak-sale period, the store service group receives a "Staff Shift Handover Note" whose original text reads: "This peak-sale shift covers November 1 to November 3, with 6 people on the morning shift and 4 on the evening shift. Temporary staff take up their posts after confirmation by the warehouse manager." A draft sentence is submitted for review: "This shift note has confirmed that all temporary staff are on duty." The original text states only that temporary staff take up their posts after confirmation by the warehouse manager; it does not state that the confirmation has been completed. The draft sentence turns "after confirmation" into "has confirmed all are on duty," which is unsupported. The retrieval action can only return the original note text; it cannot substitute for the warehouse manager's confirmation act.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model base containing shift and support plans | Cross-model bases cannot be associated together; check the vector model identifier first and exclude non-matching ones |
| Search mode | Try semantic retrieval and full-text retrieval separately first | Compare the returned content of the two modes against test questions; do not assume one mode suits all materials |
| Citation limit | Tentatively set to 2000 tokens | Test whether content stays complete using shift passages of different lengths, then compare with smaller or larger values |
| Minimum relevance | Tentatively set to 0.4 | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimization | Enable first | Compare the original question with the optimized question and check whether the subject has been altered |
| Reranking | Enable ReRank first | Confirm a rerank model exists and is enabled by setting; compare candidate order before and after reranking |
Three common errors
- Treating a retrieval return as authorized execution. Retrieving the original shift note text only shows that the text exists in the knowledge base; it does not show that temporary staff confirmation is complete. When checking, compare the retrieval output with the warehouse manager's confirmation record and observe whether the two describe the same fact.
- Writing a pending-review state as approved. When the shift note is pending review, a draft sentence saying "approved" lacks support. When checking, compare the note's current status field with the wording of the draft sentence and observe whether the same field holds mutually exclusive values.
- Using minimum relevance filtering as a substitute for permission isolation. Minimum relevance is a filtering threshold; it does not exclude unrelated business or unauthorized records. When checking, compare the returns of the same query across different knowledge base scopes and observe whether content from other scopes is mixed in.
How to verify the setup
- Input the original "Staff Shift Handover Note" above and the draft sentence "This shift note has confirmed that all temporary staff are on duty." Compare whether the original text states that confirmation is complete. The original writes only "take up their posts after confirmation by the warehouse manager" and does not state that the confirmation act is complete, so the draft sentence is judged unsupported.
- Input a short question such as "How many on the evening shift?" and compare the returns of semantic retrieval and full-text retrieval. If full-text retrieval hits "4 on the evening shift" while semantic retrieval returns a generalized shift description, record the difference between the two modes; the first result is not guaranteed to hit.
- Input a paper empty result [] and check whether the downstream node runs through the output path as documented. When the empty result has length 0, the path still runs; do not claim that a fictional rule search necessarily returns []. For cross-model bases, check the identifier first and exclude non-matching ones; do not run cross-model retrieval to compare missed recalls.