What the documents look like
Supermarket supplier entry settlement centres on purchase-and-sale or consignment payables, and a supplier does not necessarily operate a staffed counter on site. Store service and internal training therefore have to turn four material types into searchable knowledge: purchase contracts, delivery acceptance notes, rebate agreements, and statements of account. These materials are updated on a weekly or monthly rhythm, so the retrieval step has to keep track of whether a material version and its status correspond to the settlement period currently under discussion.
The main data sources are ERP purchasing, the supplier portal, POS sales, and the finance system. Once these records enter the knowledge base, retrieval only returns text fragments. It does not perform account reconciliation and does not determine authorization. In this training context, a common retrieval need is to check the stated basis for a supplier's accepted quantity in a given period, the basis for a rebate deduction, or the handling steps for a statement difference. The retrieved result needs to retain the material subject, the object, the period, and the status so that a person can continue the comparison.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases using different vector models cannot be linked together. The vector model identifier must be checked first, and any knowledge base that does not match is excluded. Semantic retrieval, full-text retrieval, and hybrid retrieval each have their own fit. Exact queries such as codes and model numbers suit full-text or hybrid retrieval, while natural-language questions suit semantic or hybrid retrieval. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not carry permission or entity isolation. A retrieval hit is not a business authorization.
Citation limit is measured in tokens, so materials of different lengths need separate observation to see whether the content is complete. Question optimization performs coreference resolution and question expansion. The original question and the optimized question need to be compared to check whether the subject was altered. An empty result is output as an array whose length can be zero, and the downstream chain still runs. An empty result does not by itself show that the material does not exist.
Hypothetical paper exercise sample: the material text for comparison reads, "Party A, the supermarket, and Party B, the supplier, signed a purchase-and-sale contract in March 2025, agreeing that the rebate deduction rate for the April 2025 settlement period is 2%. Party B's delivery acceptance note for April 2025 shows an accepted quantity of 1,200 units, and the statement of account records 60 units returned." The draft sentence to be assessed is, "The supplier's April 2025 rebate should be calculated on 1,200 units, and the return does not affect the rebate base." This sample is used only for paper comparison on this page and does not refer to any real entity or identifier.
Configuration and implementation checks
The table below is built around the sample above and lists officially documented settings or implementation checks relevant to this feature and to the store service and internal training task. The numeric values are hypothetical starting points for this exercise, used for comparison with another value, and are not stated official defaults.
| Configuration item | Hypothetical starting point for this exercise | Comparison method and reason |
|---|---|---|
| Linked knowledge bases | Link only the entry settlement material base that uses the same vector model | Check that the vector model identifier is consistent across bases; bases using different models are not linked together |
| Search mode | Start with hybrid retrieval | The sample contains both the natural-language phrase "rebate deduction" and the period limit "April 2025"; hybrid retrieval covers semantic and keyword matching at the same time |
| Citation limit | Hypothetical starting point of 800 tokens | Compare with 400 and 1,200 to observe whether the contract clause and the statement fragment are truncated |
| Minimum relevance | Hypothetical starting point of 0.35 | Takes effect only under semantic retrieval or result reranking; compare with 0.2 and 0.5 to observe whether unrelated fragments decrease |
| Question optimization | Leave off for this exercise while the sample question is complete | Compare the original question with the optimized question and check whether the subject "the supplier" was replaced by another subject |
| Result reranking | Enable according to settings when a rerank model is available | The rerank score ranges from 0 to 1 and is used for filtering; compare candidate ordering with the disabled state |
Three common errors
- Treating a retrieval hit as authorized execution. In the sample, the draft sentence concerns the calculation of the rebate base. Retrieving the contract and the acceptance note only shows that the materials were recalled. Rebate calculation and account confirmation need the finance system or a person to complete them under the authorized process, and the retrieval step does not produce an approval result.
- Linking knowledge bases that use different vector models. If knowledge bases using different vector models are linked together, that is a configuration error. The correct action is to check the vector model identifier, exclude the ones that do not match, and record the actual observation. No conclusion about missed recall or redundant results can be drawn from this.
- Treating minimum relevance as a permission or entity filter. Minimum relevance is a 0-to-1 filter threshold and does not exclude unrelated business or unauthorized records. If a rebate clause belonging to another supplier is recalled in the sample, the material subject field must be checked by a person. The similarity setting cannot be relied on for isolation.
How to verify the setup
- Enter the hypothetical sample text and draft sentence above, and under hybrid retrieval compare whether the recalled fragments contain the three pieces of information "April 2025", "rebate deduction rate 2%", and "60 units returned". The condition is met when all three appear in the cited content and the fragment subject matches the sample supplier. If the return information does not appear, record it as pending verification and do not judge the draft sentence correct.
- Change the citation limit from the hypothetical starting point of 800 tokens to 400, and compare whether the contract clause is truncated under the same query. The condition is met when the clause is complete at 800 and truncation appears at 400, and the effect of the limit on content completeness is recorded accordingly. If no truncation is observed, keep the original value and record the actual observation.
- Enter a short question containing only the period "April 2025", and compare the recalled results with question optimization enabled and disabled. The condition is met when the subject is completed after enabling and is not replaced by another supplier. If the subject changes after optimization, record it as a deviation introduced by question optimization, and that query result is not used for later comparison. For the empty-result case, a clearly stated paper input of [] can be used to check the downstream design. No claim can be made that a given rule will always return [] for a search.