What the documents look like
Omnichannel stock transfer work covers goods moving between stores and warehouses. It does not cover marketing outreach. In the store service and internal training setting, retrieval answers questions about how a document is written, what stage a request has reached, and who may act within which authorization boundary.
The everyday materials are transfer requests, stock snapshots, store shortage lists and transfer logistics notes. Their sources are ERP inventory, store POS, e-commerce orders and WMS. A transfer request names the originating store, the destination warehouse, the goods and a quantity. A stock snapshot names a statistical point in time. A shortage list names a store and goods. A logistics note names carrier and arrival information. Retrieval returns these contents as citations. It does not change the status of any document.
Because these materials carry a subject, an object, a status and an effective scope, a retrieved passage can be compared against the field it claims to describe. The comparison is only meaningful when subject, period and field match.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together. Check the model identifier first and exclude any base that does not match. Record what is observed. Do not run a cross-model retrieval to compare missed recalls.
Cited content is output as an array, and its length can be 0. When nothing is found, that output path still completes. An empty result can be checked against a clearly stated paper input of [] to examine downstream design. Do not claim that a retrieval over an invented rule will return [].
Retrieval is not authorization. Forms collect values. Business verification, approval, database writes and external synchronization each need their own stated downstream implementation. A pending review does not mean an operator lacks permission. Signing, reaching an agreement and a contract taking effect are three separate facts. A quotation validity period does not state a supply period.
The overall retrieval flow combines several capabilities, and the optional paths in the documented flow are not all executed on every request. Image retrieval requires an image input and a model that supports it. Rerank requires a rerank model and must be enabled in the settings. RRF is a fusion step after hybrid retrieval, not a switch.
Hypothetical drill sample. The material text to be checked reads: "Transfer request A-07: originating store is Sample Store A, destination warehouse is Sample Warehouse B, goods are Sample Model X, quantity is 30 pieces, request date is Sample Date D1, current status is pending review." "Stock snapshot S-03: statistical point in time is Sample Date D1, available quantity of Sample Model X at Sample Warehouse B is 12 pieces." The draft sentence to be judged reads: "Sample Store A has been approved to move 30 pieces of Sample Model X into Sample Warehouse B, and the goods can be put on the shelf immediately on arrival." The draft sentence treats pending review as approved, treats the requested quantity as an executed quantity, and adds a shelf-placement condition that the source text does not state.
Configuration and implementation checks
The table below lists only documented settings or implementation checks relevant to this page's materials. The values are starting points for this hypothetical drill, used for comparison against the sample source text. They are not official defaults.
| Configuration item | Starting point for this hypothetical drill and reason | What it is compared against |
|---|---|---|
| Associated knowledge bases | Associate only bases containing transfer requests, stock snapshots, shortage lists and logistics notes, and whose vector model identifiers agree | Whether the vector model identifiers of the bases are the same |
| Search mode | Start with hybrid retrieval, then compare full-text retrieval separately for numbering and model-number questions | Cited content for the same question under semantic, full-text and hybrid modes |
| Citation limit | Start at a token amount that can hold the body of one complete transfer request, because that material must be checked as a whole | Whether cited content is truncated and whether it contains the complete status field |
| Minimum relevance | First annotate relevant and irrelevant passages in the sample, then take a value that keeps irrelevant passages out of the citations | The annotation result and the filtered citation list |
| Question optimization | Enable for follow-ups such as "what is the second point", and compare before and after for complete questions that already contain subject and object | Whether the subject of the original question is altered in the optimized question |
| Result rerank | Enable when candidates are numerous, and filter using the rerank score | The rerank score and the citation order |
Full-text retrieval uses BM25 when the vector store is Milvus (≥ 2.5.16), with language identification defaulting to lingua. Other vector stores use MongoDB $text with jieba. Rerank produces a score from 0 to 1. Minimum relevance takes effect only under semantic retrieval or when result rerank is used. Question optimization adds a model call before retrieval and therefore usually increases overall latency.
Three common errors
- Reading a pending-review status as approved. The sample source text states "current status is pending review", while the draft sentence writes "has been approved". This is a contradiction between mutually exclusive values under the same subject, the same period and the same field. The input is the source text and the draft sentence; the comparison is the value of the status field; the observable result is that the two do not agree.
- Reading a requested quantity as an executed quantity. The source text gives "quantity is 30 pieces" as the requested amount, and the snapshot gives an available quantity of 12 pieces. The draft sentence writes that 30 pieces are being moved in and says the goods can be put on the shelf on arrival. The input is the two source texts and the draft sentence; the comparison is the quantity field and whether any execution or shelf-placement condition is stated; the observable result is that the draft sentence introduces an action the source text does not state.
- Treating a retrieval result as an authorization record. Retrieval returns cited content only and does not verify approval. The input is one retrieval request; the comparison is the returned citations against the authorization status field; the observable result is that when the citations contain no approval conclusion, the request remains to be verified.
How to verify the setup
- Input the sample source text and the draft sentence, compare the value of the "current status" field, and observe the result: when the draft sentence disagrees with the source text, record it as a contradiction and do not record it as approved.
- Input the sample source text and the draft sentence, compare the requested quantity, the available quantity in the snapshot and the action described in the draft sentence, and observe the result: when the draft sentence adds an execution or shelf-placement condition the source text does not state, record it as unsupported.
- Input a transfer question that contains an explicit subject and object, compare the retrieval terms and cited content before and after question optimization is enabled, and observe the result: when the subject is altered or the referent shifts, record that observation. Do not conclude from it that retrieval necessarily fails.
Where the material does not state a field, or where only evidence of a related stage exists, record the item as unsupported or to be verified, and keep the evidence gap open. Do not supply an authorization or effective clause to obtain a definite answer.