What the documents look like
Multi-warehouse allocation and inventory synchronisation material covers allocation rules, inventory synchronisation tables, transfer orders and stockout alerts. Sources include WMS, OMS, ERP inventory and platform inventory interfaces, with update frequency described as daily or hourly. In store service and internal training, supply chain planners, WMS administrators and e-commerce operators retrieve these documents to answer store questions about dispatch warehouse priority, sellable inventory definitions, transfer status and stockout alert handling. Retrieval output supports training explanation and store replies. Retrieval itself is not authorisation, and it does not replace inventory verification, approval or database writes.
The 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; the model identifier must be checked and non-matching bases excluded. Reference content is output as an array and its length can be 0, meaning the downstream path still runs when nothing is retrieved, so downstream design must handle empty results. Search mode must be compared across semantic retrieval, full-text retrieval and hybrid retrieval for the test question, without assuming one mode suits all material. The reference upper limit uses tokens as its unit, and content completeness should be tested with material of differing lengths. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, so relevant and irrelevant samples should be labelled before adjusting the threshold. Question optimisation performs coreference resolution and question expansion, and the original question should be compared with the optimised question to check whether the subject was altered. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled by configuration. Retrieval and relevance do not provide permission or entity isolation.
Hypothetical paper exercise sample: the material text for checking reads "East China warehouse and South China warehouse allocation rule: orders for the East China region are dispatched with priority from the East China warehouse; orders for the South China region are dispatched with priority from the South China warehouse; transfer order DB-0417 has status pending review and the transfer is not yet complete; the sellable inventory synchronisation table updates daily and yesterday's synchronisation batch has been generated." The draft sentence to be assessed reads "Transfer order DB-0417 has completed the transfer, and the store may dispatch directly against South China warehouse sellable inventory." This sentence conflicts with the source status "pending review" and is judged unsupported.
Constraints at the retrieval step
The association rule limits a single search node to knowledge bases sharing one vector model. This is a configuration boundary, not a statement about retrieval quality. When two bases use different vector models, the correct action is to check the identifier, exclude the non-matching base and record the observation. No runtime outcome such as missed recall or redundant returns should be asserted for an association that is not permitted.
The reference content array can be empty, and the output path still completes. Downstream branches therefore need an explicit empty-result design. A paper input of an empty array can be used to check whether that branch exists; it does not demonstrate that a particular query returns an empty array.
Search mode selection is a comparison task. Semantic retrieval suits natural-language questions and semantically close content. Full-text retrieval suits identifiers, model numbers, proper nouns and error codes. Hybrid retrieval uses both and merges results through RRF, which is a fusion process rather than a switch. Reranking applies a model to reorder candidates and produces a score between 0 and 1, which can be used for filtering. The minimum relevance value applies only under semantic retrieval or reranking. Question optimisation adds a model call before retrieval and can improve multi-turn retrieval while increasing elapsed time.
Image retrieval is an optional path in the overall flow. It requires image input and the relevant model capability, and whether an image can be retrieved also depends on whether a corresponding index was created at ingestion. Reranking likewise requires a model and must be enabled by configuration. The presence of these paths in the documented flow does not mean every request executes them.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the allocation rule base that uses the same vector model | Check the vector model identifier of each base, exclude non-matching bases, and do not run cross-model retrieval |
| Search mode | Run one paper comparison each for semantic retrieval and full-text retrieval | Record hits for the sample question under each mode and compare identifier-type and natural-language-type questions |
| Reference upper limit | Start from a value that covers one complete allocation rule | Compare whether reference content is truncated using rule texts of different lengths, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set a threshold | Takes effect only under semantic retrieval or result reranking; observe whether labelled samples are retained after filtering |
| Question optimisation | Compare the original question with the optimised question for a follow-up containing a reference | Check whether the optimised subject remains the allocation rule; record when the subject is altered |
| Result reranking | Compare rerank scores when a model is available and the setting is enabled | Observe the 0-1 score and candidate ordering; do not treat the score as a permission filter |
Three common errors
- Treating a pending-review transfer order as a completed transfer. The sample source states that DB-0417 has status pending review, while the draft sentence writes that the transfer is complete. Input is the source status field and the draft sentence, comparison is the value of the same field on the same document, and the observable result is a value mismatch, judged unsupported.
- Treating a retrieval hit as dispatch authorisation. The retrieved allocation rule states only the priority dispatch warehouse and does not record authorisation for a store to dispatch directly. Input is the rule text and the draft request, comparison is whether an authorisation statement exists, and the observable result is that the source records no such authorisation, judged pending verification.
- Treating minimum relevance as an entity or permission filter. With a hypothetical input query for "South China warehouse", the observable result may include content from other regions that contains the same wording. The comparison is whether filtering by warehouse entity occurs, and the observable result is that no entity isolation is applied, so business verification requires a separate system implementation.
How to verify the setup
- Input the sample source text and the draft sentence "transfer order DB-0417 has completed the transfer", compare the status value of the same field on the same document, and observe that the draft sentence conflicts with the source value "pending review". The condition is met when the result is judged unsupported and the evidence gap is recorded.
- Input the sample source text and the draft request "the store may dispatch directly against South China warehouse sellable inventory", compare whether the source records dispatch authorisation, and observe that the source states only a priority dispatch warehouse. The condition is met when the result is judged pending verification without supplying an authorisation clause.
- Input a paper empty result of an empty array, check whether the downstream path continues with a reference length of 0, and observe whether the downstream design includes an empty-result branch. The condition is met when the empty-result handling design is recorded, without claiming that a fictional rule query returns an empty array.