What the documents look like
Supermarket supplier entry and settlement materials centre on four document types: purchase contracts, delivery acceptance notes, rebate agreements and reconciliation statements. Data originates in ERP purchasing, supplier portals, POS sales and finance systems. Compared with concession counters in hypermarkets, this category usually involves purchase-and-sale or consignment payables, and the supplier does not necessarily operate on site. Errors in this area produce payment disputes, supplier suspension of deliveries, and exposure to fair-trading requirements. Product information and selection questions typically ask what state a given item occupies across the contract, acceptance, rebate and reconciliation materials, which fields corroborate each other, and which fields appear in only one document type. Materials update weekly or monthly. The retrieval step answers from the materials that exist; it introduces no batch synchronisation and no real-time trigger.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Before retrieval, check the knowledge base identifier and the vector model identifier, exclude those that do not match, and then run the search. Do not use cross-model retrieval to compare missed recalls. Hybrid retrieval fuses multiple result paths through RRF; RRF is a fusion process, not an independent switch that can be turned on. Result reranking requires a rerank model and must be enabled through settings. Image retrieval requires corresponding image input and model capability. These paths do not execute on every request. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The citation limit is measured in tokens. Semantic recall and relevance scores carry no permission or entity-isolation duty, and a retrieval result does not constitute authorisation.
Hypothetical exercise sample: a supplier business staff member submits a paper document set for verification of "biscuit category goods at one East China store, March 2025". The source text reads: "The purchase contract states that payment terms for this item are 30 days after delivery; the delivery acceptance note dated 12 March 2025 records 480 cases actually received; the rebate agreement states a quarterly deduction rate of 2%; the March 2025 reconciliation statement lists the payable amount and the net amount after deduction." The draft sentence to be assessed is: "Settlement for this item in March 2025 has been completed." The subject, period and fields in that sentence must be compared item by item against the materials above.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model base containing the four document types above | Check the base identifier and the vector model identifier first; exclude bases on different models |
| Search mode | Run semantic retrieval and full-text retrieval separately | Compare recalled content for the same question under both modes; assume no mode is superior in advance |
| Citation limit | Hypothetical starting point of 3000 tokens | Test with material fragments of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then compare remaining content after filtering |
| Question optimisation | Not enabled for this sample, whose question is complete | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Not enabled for this exercise | Requires a rerank model enabled through settings; this exercise does not assume one is available |
Three common errors
- Treating a retrieval hit as established authorisation. Input is the draft sentence above together with the retrieval results. Compare whether the retrieved content contains an approval or effective record required for completed settlement. If the materials only list contract payment terms, accepted quantity, deduction rate and reconciliation amount, the settlement state remains unsupported.
- Associating knowledge bases that use different vector models. Input is two base identifiers and their respective vector model identifiers. Compare whether the two are identical. If they are not, exclude them and do not run cross-model retrieval.
- Treating a quotation validity period or a payment term as a supply period or as completed settlement. Input is the contract payment term clause and the reconciliation statement. Compare whether the field names and values point to the same fact. If the fields differ, neither can be inferred from the other.
How to verify the setup
- Input the four source document texts from the hypothetical sample and the draft sentence. Compare whether the subject, period and fields in the draft sentence have corresponding records in the materials. The sentence is supported only when the subject, period and field match and the value agrees. The materials here contain no record of completed settlement, so the result should be unsupported.
- Input two knowledge base identifiers whose vector model identifiers differ. Compare whether the identifiers match. If they do not, record them as not associable together and exclude them, and do not continue retrieval.
- Input a paper citation array explicitly set to empty. Compare whether the downstream design continues along the empty-citation path. If the downstream design explicitly accepts an empty array, the condition is met. Do not claim that a fictional rule retrieval necessarily returns an empty array.