What the documents look like
Supermarket supplier entry and settlement work runs on purchase contracts, delivery acceptance notes, rebate agreements and reconciliation statements. The data comes from ERP purchasing, supplier portals, POS sales and finance systems, with weekly or monthly update cycles. The people involved are supermarket buyers, supplier business staff and finance settlement staff. Compared with large-format joint-operation counters, suppliers in this category settle mainly through purchase-and-sale or consignment arrangements and do not necessarily operate on site. Product information and selection Q&A therefore has to cover contract terms, accepted quantities, rebate deduction rates and return differences at the same time.
The retrieval target for product information and selection Q&A is the product entries, specifications, settlement methods and agreement terms inside those documents. The vector model and index determine whether this content can be recalled semantically, and whether exact queries such as item numbers and model numbers can be matched. The knowledge base used for retrieval must share the same vector model, so the actual model and its input type are checked first. One source data block can be associated with multiple indexes, and the index content is checked separately from the source text. Index content is best kept complete and focused, with retrieval coverage weighed against index length using samples. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion time, and cannot be judged from the query side alone. After a model or index change, the same question set is used to compare the cited source text, and vector scores alone do not support a promise about outcomes.
Constraints at the vector indexes step
A hypothetical paper exercise sample, not a real record:
Material A (excerpt from a purchase contract): Party A is a supermarket, Party B is a food supplier, the contract period runs from January 2025 to December 2025, the settlement method is purchase and sale, and the payment term is 30 days.
Material B (excerpt from a rebate agreement): After Party B's annual supply amount reaches the agreed tier, Party A calculates the rebate at the corresponding deduction rate, and the deduction rate is based on the figure confirmed in the reconciliation statement.
Draft sentence to be assessed: "The supplier's 2025 annual rebate deduction rate has been executed at the highest tier."
Handover request to be assessed: "Please confirm this supplier's rebate tier for the current year and synchronize it to the settlement statement."
Material B states only that the deduction rate is based on the figure confirmed in the reconciliation statement. It does not state the tier value or the execution result, so the draft sentence and the handover request both remain pending verification.
The same vector model requirement applies to the knowledge bases being associated for retrieval. It does not extend to business registration sites or other external systems. Semantic recall and relevance do not carry authorization or entity isolation. Retrieval is not authorization.
Configuration and implementation checks
| Configuration item | Value or check item on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Purchase contracts, delivery acceptance notes, rebate agreements, reconciliation statements | Only knowledge bases with the same vector model can be associated together |
| Vector model | Check the actual model identifier and input type first | Different models require separate knowledge bases |
| Index content | Product entries, specifications, settlement methods, agreement terms | Complete and focused, with length weighed against coverage |
| Image index | Check whether ingestion generated image descriptions or image vectors | Uploading an image at query time does not replace an ingestion index |
| Search mode | Semantic search, full-text search or hybrid search | Item number and model number queries lean toward full-text search |
| Result review | Compare cited source text using the same question set | Vector scores do not support an outcome promise |
For this exercise, the starting values for the citation limit and the minimum relevance are each set to a value to be recorded. The reason for choosing them is to make before-and-after comparison easy on the same question set, and the comparison object is the cited source text before and after adjustment. These are not official default values.
Three common errors
- Associating knowledge bases that use different vector models. The check input is the model identifier of each knowledge base, the comparison object is whether the identifiers match, and the observable result is that a knowledge base is excluded and recorded when its identifier does not match. Cross-model retrieval is not executed to compare missed recalls.
- Treating an image uploaded at query time as an existing image index. The check input is whether ingestion generated an image description or image vector, the comparison object is the query-side input, and the observable result is that the gap is recorded when ingestion produced no index.
- Treating the draft sentence under assessment as a confirmed conclusion. The check input is the source text of Material B, the comparison object is the tier and execution wording in the draft sentence, and the observable result is that the item is recorded as pending verification when the source text does not state the tier value.
How to verify the setup
- For Material A and Material B, input the draft sentence "The supplier's 2025 annual rebate deduction rate has been executed at the highest tier." Compare whether Material B states the tier value and the execution result. The condition is met when the source text does not state them, and the draft sentence is judged unsupported.
- For the handover request, input the request text. Compare whether the confirmation basis in Material B is the figure confirmed in the reconciliation statement. The condition is met when the request lacks a source for that confirmed figure, and the item is recorded as an evidence gap.
- Use the same question set to compare the cited source text before and after adjustment. The input is the same question set, the comparison object is whether the cited source text changes, and the observable result is that differences are recorded when it changes. An empty result is checked against the downstream design using a paper input of an empty array.
Retrieval or a form does not constitute authorized execution. Confirming a rebate tier and synchronizing a settlement statement require the settlement process to be carried out separately.