What the documents look like
Supermarket supplier entry settlement centres on purchase-and-sale or consignment payables, and the supplier does not necessarily operate a staffed counter on site. That difference shapes the handover scope directly. Four materials form the core record set: the procurement contract, the delivery acceptance note, the rebate agreement, and the statement of account. The main data sources are ERP purchasing, the supplier portal, POS sales, and the finance system. Materials update weekly and monthly, and the retrieval layer does not infer batch synchronisation or real-time triggers from that cadence.
The retrieval task in operations permission and handover management is to surface the subject, object, status, and validity scope recorded in those materials so a person can check them. Retrieval returns quoted content only; it does not perform authorisation decisions. Quoted content is output as an array whose length may be zero, and the output path still completes when nothing is retrieved.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; in that situation the check is limited to comparing vector model identifiers, excluding non-matching bases, and recording what was observed, without running a cross-model retrieval to compare missed recalls.
Search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval for a given test question, and no single mode is presumed suitable for every material. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch. Result reranking requires a rerank model and must be enabled in settings, and image retrieval requires corresponding image input and model capability. Optional paths shown in the overall flow diagram do not mean every request executes all of them.
Minimum relevance is a 0–1 filter threshold that takes effect only under semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation. The citation limit is measured in tokens and controls the maximum amount of content cited per search. Question optimisation performs coreference resolution and question expansion, and it adds one model call.
Configuration and implementation checks
Hypothetical drill sample: a supermarket signs a purchase-and-sale procurement contract with a supplier. The contract states a settlement cycle of monthly payment within 30 days. The rebate agreement provides for a quarterly rebate calculated at 2% of purchase value. The delivery acceptance note records received quantity and discrepancies, and the statement of account lists the amount payable for the period. A draft handover sentence reads: "The rebate deduction point for supplier entry settlement follows the contract, and the successor continues to check it after handover." That sentence does not state the rebate rate, the calculation period, or the authorisation boundary, so it is unsupported.
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Procurement contract base, rebate agreement base | Both may be associated only when their vector model identifiers match; exclude a base when they differ |
| Search mode | Hybrid retrieval | Run the same test question under semantic, full-text, and hybrid retrieval, then compare quoted content |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with delivery acceptance notes and statements of account of differing length, check content completeness, then adjust |
| Minimum relevance | Hypothetical starting point of 0.5 | Label related and unrelated samples first, then compare quoted content after filtering |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled according to settings | Where a rerank model is available, compare candidate ordering before and after reranking |
Three common errors
- Treating retrieval output as an authorisation action. The input is the draft handover sentence together with the quoted content returned by retrieval; the comparison is whether the quoted content states an authorising subject and a validity scope; the observable result is that when the quoted content states no authorisation boundary, the request remains pending verification.
- Treating minimum relevance as a permission or entity filter. The input is the minimum relevance setting and statements of account for different subjects; the comparison is the returned content before and after filtering; the observable result is that the relevance threshold does not exclude records belonging to unrelated subjects.
- Forcing knowledge bases on different vector models into one association. The input is the vector model identifiers of two knowledge bases; the comparison is whether those identifiers match; the observable result is that a base is excluded when the identifiers differ, and no cross-model retrieval is run for comparison.
How to verify the setup
- Check one: input the procurement contract text from the hypothetical sample and the draft sentence "the rebate deduction point follows the contract"; compare whether the draft sentence states the rebate rate and the calculation period; the observable result is that when the source text does not state that rate, the sentence is judged unsupported and the evidence gap is retained.
- Check two: input the delivery acceptance note and the statement of account from the hypothetical sample; compare the received quantity field and the amount payable field for the same supplier in the same period; the observable result is that mutually exclusive values in the same field indicate a contradiction, while evidence covering only a related stage indicates a pending verification.
- Check three: input an explicit paper empty result []; check whether the downstream design continues along the output path when the citation length is zero; the observable result is that the condition is met when the downstream does not break on the empty array. Retrieval checks record observations and failure branches only, and do not guarantee a first hit or full recall.