What the documents look like
Brand omnichannel stock transfer work covers goods moving between stores and warehouses. It does not cover marketing outreach. The main data sources are ERP inventory, store POS, e-commerce orders and WMS. The daily documents are transfer requests, stock snapshots, store shortage lists and transfer logistics notes. Operating reports and exception reviews retrieve and recall against these documents.
A transfer request records the sending party, the receiving party, the goods and the quantity. A stock snapshot records book available quantity at a point in time. A store shortage list records missing goods and demand. A transfer logistics note records dispatch, in-transit and arrival. These documents update daily or weekly, and that update frequency does not change the retrieval and recall conditions on this page. Errors lead to overselling, store stockouts and book-versus-physical mismatch, so a retrieval result must point to a specific document and field for human checking.
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 using different vector models cannot be associated together. Check the vector model identifier first, exclude non-matching bases, and record what is observed. The smallest search unit is the knowledge base; a collection only classifies data and does not affect search results.
Semantic retrieval depends on vector distance. Full-text retrieval depends on keyword hits. Hybrid retrieval uses both and merges them through RRF. RRF is a fusion process, not a switch. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. It does not carry permission or entity isolation. The reference limit is measured in tokens and controls how much content one search may cite. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and must be enabled by configuration. When the reference output is empty, its length can be 0 and the downstream path still runs; a paper input of [] can be used to check the downstream design. Retrieval and recall do not constitute authorization; approval, database writes and synchronization need separately stated downstream implementation.
Configuration and implementation checks
Hypothetical drill sample: on 10 March 2025 a brand retail supply chain team reviews a transfer request whose text reads "Sending party: East China warehouse; receiving party: Hangzhou Hubin store; goods: SKU-8842; quantity: 120 units; request status is pending review." A same-day stock snapshot reads "East China warehouse SKU-8842 book available quantity: 300 units." A draft sentence to be judged reads "This transfer request has been approved and inventory has been deducted."
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base holding transfer requests and stock snapshots | Check the vector model identifier first, exclude different-model bases, then compare whether recalled content falls inside that base |
| Search mode | Hybrid retrieval | Number and status fields need full-text hits; natural-language description needs semantic hits; compare whether both recall paths appear |
| Reference limit | Start from covering one transfer request plus one stock snapshot | Test with materials of different lengths to see whether content is complete, then adjust the limit |
| Minimum relevance | Mark relevant and irrelevant samples first, then set a value | This value takes effect only under semantic retrieval or reranking; compare whether required fields survive filtering |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether subject and period were altered |
| Result reranking | Enable when candidates are many | Rerank scores run from 0 to 1; compare whether documents carrying the status field move earlier |
Three common errors
- Treating a retrieval result as an authorization conclusion. In the sample the request status is "pending review," while the draft sentence says "has been approved." The input is the request text and the draft sentence; compare the status fields; the observed result is that the draft sentence goes beyond what the text records. Pending review does not show a lack of operating permission, and it is not approval.
- Associating knowledge bases that use different vector models. The input is the vector model identifiers of the two bases; compare whether the identifiers match; the observed result is that different identifiers cannot be associated together. Do not invent runtime outcomes such as missed recall or redundant returns; record only the identifier mismatch and the exclusion.
- Treating minimum relevance as a permission or entity filter. The input is a candidate set containing unrelated business fragments; compare whether unrelated fragments remain before and after filtering; the observed result is that a relevance threshold only affects the acceptance of low-relevance content and does not exclude unauthorized records.
How to verify the setup
- For "Sending party: East China warehouse; receiving party: Hangzhou Hubin store; goods: SKU-8842; quantity: 120 units; request status is pending review" together with the same-day snapshot "East China warehouse SKU-8842 book available quantity: 300 units," enter the search question "SKU-8842 transfer request status." Compare whether the recalled content contains both the request status field and the snapshot quantity field. The observed result is that both source texts are cited and the status field shows pending review. If only one is recalled, record the gap and check chunking and search mode.
- For the draft sentence "This transfer request has been approved and inventory has been deducted," the input is the request text and the draft sentence. Compare the status field and the quantity field. The observed result is that the draft sentence disagrees with the recorded status, and no inventory deduction is recorded in the source text, so the sentence is unsupported. Do not supply authorization or effectiveness clauses to obtain a definite answer.
- For an empty reference output, the input is the paper input []. Compare whether the downstream path continues to run. The observed result is that the path runs and the reference length is 0. Do not claim that a fabricated rule search necessarily returns []; record only the design check under that paper input.