What the documents look like
Omnichannel stock transfer work follows goods as they move between stores and warehouses. The documents that carry this work are transfer requests, stock snapshots, store shortage lists, and transfer logistics notes. This material does not cover marketing outreach, which separates it from membership and private-domain operations. Retrieval questions instead centre on available stock, transfer routing, and arrival or receipt status.
Stock snapshots and shortage lists are described as updating daily or weekly. That stated frequency does not by itself establish real-time synchronisation, batch synchronisation, or transaction-triggered refresh, and this page does not infer any of those. When a question depends on a date, the date range written on the document governs.
Product data and selection questions usually contain store names, product codes, specifications, and transfer order numbers. Questions built around codes and order numbers lean on exact keyword hits, while natural-language questions lean on semantic recall. Both paths can be present in one knowledge base search.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that share the same vector model. Knowledge bases with different vector model identifiers cannot be associated together; the identifiers are checked first and non-matching bases are excluded. That same-model requirement applies to the knowledge bases selected inside the node. It does not extend to external systems such as business registration sites, and it says nothing about whether those systems use the same model.
Retrieval is a recall action. It does not perform authorisation checks. A recalled transfer request is not evidence that the request was approved, and the same-model association rule does not turn retrieval into a permission filter. A relevance score is a similarity measure, not a category, entity, or permission filter, and it does not exclude unrelated businesses or unauthorised records.
Reference content is output as an array, and its length can be zero. The output path still runs when nothing is retrieved. The minimum relevance setting is a filter threshold that applies only under semantic retrieval or when result reranking is used; it is not a final sort rule. Question optimisation performs coreference removal and question expansion and adds an extra model call before retrieval. RRF is the fusion step that combines results from multiple recall paths; it is not an on/off switch. Image retrieval requires an image input plus the relevant model capability, and reranking requires a rerank model that is enabled in the settings. The overall retrieval flow lists optional paths, and not every request exercises all of them.
Configuration and implementation checks
Hypothetical drill sample: the source text reads, "Transfer request A-0712, requesting party a store in the East China region, object product code SKU-3301, quantity 12 pieces, request date 12 July 2026, status pending review." The draft sentence under assessment reads, "Transfer request A-0712 has been approved and the transfer can be executed." The source text states only pending review and does not state approval.
| Configuration item | Hypothetical starting point for this drill | Comparison and notes |
|---|---|---|
| Associated knowledge bases | One base whose vector model identifier matches | Check each base's vector model identifier first and exclude non-matching ones |
| Search mode | Start with semantic retrieval for the first pass | Compare results item by item against full-text and hybrid retrieval |
| Reference limit | Set to 2000 tokens as a drill starting value | Chosen because the sample carries an order number and a status field; compare against 1000 and 4000 for content completeness |
| Minimum relevance | No threshold set at first | Label sample content as relevant or irrelevant first, then compare filter behaviour |
| Question optimisation | Off in the first round | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Not enabled initially | Compare ordering only after a rerank model is available and enabled in settings |
The reference limit is expressed in tokens. The starting value above is a drill choice, not an official default. The comparison method is to run the same labelled questions at 1000, 2000, and 4000 tokens and observe whether the order number, status field, and date range survive into the reference content.
Three common errors
- Treating a pending-review status as approved. Input the draft sentence and the source text, compare the status field for the same request number A-0712, and note that the source states pending review, so the draft sentence is not supported.
- Treating retrieval recall as authorisation to execute. Input a recalled transfer request, compare its status field and any approval record, and note that recall alone does not constitute approval; execution requires a separate downstream business system implementation.
- Associating knowledge bases with different vector model identifiers. Input two bases with different vector model identifiers, compare the identifier fields, and exclude the non-matching base rather than running a cross-model search to compare missed recalls.
How to verify the setup
- Input the source text and the draft sentence, compare the status field for request number A-0712, and treat the draft sentence as unsupported when the source states pending review while the draft claims approval.
- Input a natural-language question drawn from a store shortage list, compare the reference content returned by semantic retrieval against full-text retrieval, and observe whether the code field is hit by full text and whether the semantic result is relevant; record a gap when the code is not hit.
- Input an explicit paper empty result of [] and compare how the downstream node is designed to handle it, observing whether the empty-reference path runs through. Do not claim that any rule-based search necessarily returns an empty array.
Across all three checks, the comparison must use the same subject, the same period, and the same field. A status field with explicitly mutually exclusive values is the only case that counts as a contradiction. A field that is not stated, or evidence that only covers a related stage, is unsupported or pending verification. Pending review does not by itself show a lack of operating permission, and signing, reaching agreement, and contract effectiveness are three separate facts. Where a required document is missing from the accessible retrieval scope or from a clearly identified manual input, record the gap rather than supplying an assumed authorisation or effective clause.