What the documents look like
Omnichannel stock transfer work covers goods moving between stores and warehouses. The main data sources are ERP inventory, store POS, e-commerce orders and WMS. The documents handled day to day are transfer requests, inventory snapshots, store shortage lists and transfer logistics notes. These materials update daily or weekly, and retrieval results reflect the state recorded when the material was ingested. A retrieval result does not replace transfer approval, inventory deduction or receiving into stock.
Operations permission and handover management concerns the authorization state of a given material, action or request at handover time. Knowledge base retrieval and recall can only return reference content. Reference content does not constitute authorization and does not constitute approval. The retrieval step has to keep "found the relevant material" separate from "is permitted to execute the transfer."
The three most time-consuming steps in this category are locating transferable stock, approving the transfer route, and updating receiving into stock. None of these steps is performed by retrieval. Retrieval supplies the text an operator reads during handover, and the operator or a downstream system decides what the text supports.
Constraints at the retrieval step
A knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Before retrieval, the vector model identifier has to be checked, non-conforming bases excluded, and the actual observation recorded. Cross-model retrieval is not executed to compare missed recalls.
When the vector store is Milvus (≥ 2.5.16), full-text retrieval uses BM25, with vectors and full text stored in the modeldata_v2 single table. Other vector stores (PG, OceanBase, SeekDB, openGauss) use MongoDB $text with jieba tokenization. Hybrid retrieval merges multiple recall paths through RRF. ReRank performs a second ordering. The final result passes through similarity filtering and a reference upper limit.
The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. That threshold is a filtering condition. It does not perform permission or entity isolation. A similarity score does not distinguish store, warehouse or brand, and retrieval does not automatically exclude a record because it belongs to another entity. The reference upper limit is measured in tokens and controls how much content a single search may reference. The knowledge base search node outputs references as an array, and the length can be 0, which means the output chain still runs even when nothing is retrieved.
Hypothetical paper exercise sample: in the handover records of a brand retail supply chain, a passage of material available for checking reads: "Transfer request T-2024-0312 was submitted by Store A on 12 March 2024, requesting 20 units of a product be transferred from Store A to e-commerce warehouse B, with status pending review. Inventory snapshot S-20240312 shows Store A has 35 units of that product available. Store shortage list Q-20240312 lists e-commerce warehouse B as short 15 units of that product." A separate draft sentence reads: "This transfer request has been approved and the transfer may be executed." The point to determine is whether the phrase "has been approved" in the draft sentence has material support.
Configuration and implementation checks
The configuration below centers on the sample above. It lists documented settings or implementation checks relevant to this feature and this task. The values are hypothetical starting points for this exercise, used for comparison, and are not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the transfer material base that uses the same vector model | Check the vector model identifier first, exclude bases with a different model, and record the actual association result |
| Search mode | Hybrid retrieval | The sample contains both the identifier T-2024-0312 and natural-language description; compare semantic, full-text and hybrid modes on the same question |
| Reference upper limit | 800 tokens | The sample material is about three sentences; test with longer and shorter transfer logistics notes for content completeness, then adjust |
| Minimum relevance | 0.3 | Label relevant and irrelevant samples first, compare reference content under different thresholds; this value takes effect only in semantic retrieval or reranking |
| Question optimization | Enabled | Compare the original question with the optimized question, checking whether subject, document number and date were altered |
| Result reranking | Enable ReRank | When candidates are numerous, compare ordering before and after reranking; reranking requires a model and must be enabled per settings |
Three common errors
- Treating a retrieval hit as approval. Input the draft sentence "This transfer request has been approved," and retrieve transfer request T-2024-0312, whose status field reads "pending review." Compare the same document number, the same period and the same status field. Pending review and approved are mutually exclusive values, so the draft sentence has no material support. Retrieving the request form does not mean approval is complete.
- Associating knowledge bases with different vector models. Input the vector model identifiers of two knowledge bases and compare whether they match. When the identifiers differ, exclude the non-conforming base and record the actual observation. Cross-model retrieval is not executed to compare missed recalls, and the runtime result of a combined association is not inferred.
- Treating the relevance threshold as a permission filter. Input a transfer request belonging to another store and compare whether the retrieval result is excluded because of the minimum relevance setting. Relevance is a similarity filtering condition and does not exclude records by entity or permission. Retrieving that record does not mean the current operator is permitted to process it.
How to verify the setup
- Verify the draft sentence determination. Input the complete sample material text and the draft sentence "This transfer request has been approved and the transfer may be executed," then compare the status field "pending review" of transfer request T-2024-0312 with the phrase "has been approved" in the draft sentence. When the same document number, the same period and the same field carry mutually exclusive values, the draft sentence has no support. If the material states only "pending review," the determination is pending verification, not a negation or a conflict.
- Verify empty-result handling. Input a clearly paper-based retrieval result of
[]and check whether the downstream chain runs. The knowledge base search node outputs references as an array whose length can be 0, and the output chain still runs when nothing is retrieved. Do not claim that a hypothetical rule-based retrieval necessarily returns[]. - Verify same-model association checking. Input the vector model identifiers of two knowledge bases and compare whether they match. When the identifiers match, record the association check as passing; when they differ, exclude the non-conforming base and record the gap. Cross-model retrieval is not executed to compare missed recalls, and the runtime result of a combined association is not inferred.
Retrieval and recall do not carry authorization decisions, approval execution, inventory deduction or receiving into stock. A pending-review status, a quotation validity period and a signature status in handover material are different facts. Where a fact is not stated, or only related-stage evidence exists, it remains unsupported or pending verification.