What the documents look like
Price and display audits take place during ordinary trading periods, where the emphasis falls on sustained execution of standards. The working documents are store-visit photographs, price-tag lists, display standards and rectification notices, with weekly and monthly update frequencies. Copy drafting draws on these materials, and the version and effective scope of a material directly determine which sentence it can support. Within one store, the price-tag list, the current promotion calendar and the store-visit photographs are separate sources, and these sources do not automatically corroborate one another.
Constraints at the citation tracing step
A knowledge-base search node can associate one or more knowledge bases that share the same vector model. Cited content is output as an array, and its length may be zero, so the empty-result path still completes. Once the citation array reaches the answering stage, each answer sentence is checked against the cited source text. A similarity score serves relevance ordering only and does not perform store attribution, permission control or entity isolation. Hybrid retrieval is merged through RRF, reranking requires a rerank model and must be enabled by configuration, and image retrieval requires image input together with model capability; these paths do not indicate that every request executes all of them.
Hypothetical paper exercise sample. Source text one: "March 2025 East China store-visit record: store A-012 fresh-produce price tags match the POS price list, and the display conforms to the standard in force." Source text two: "March 2025 promotion calendar notice: store A-012 participates in the spring spend-and-save promotion, running from 1 March to 31 March." Draft sentence: "Store A-012 executed the spring spend-and-save promotion strictly, and its price tags matched the system price list over the long term." Determination to be made: the phrase "over the long term" in this sentence is supported only by a single March store-visit record, while the calendar notice states participation and start and end dates without stating execution results. The sentence exceeds the scope of the materials and is unsupported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Only the price-tag list base and the store-visit record base are selected | They can be associated together only when their vector model identifiers match; check the identifiers one by one and exclude any that do not match |
| Search mode | Hybrid retrieval | Store numbers and calendar names are exact terms; merge them with semantic recall and then inspect the candidates |
| Citation limit | Hypothetical starting point of 2000 tokens for this exercise | Chosen because the sample materials are short; compare against the actual candidate length before adjusting, and this is not an official default |
| Minimum relevance | Hypothetical starting point of 0.3 for this exercise | Used to filter low-relevance candidates; compare the candidate count before and after filtering, and this is not an official default |
| Result reranking | Enabled when a rerank model is available | When enabled, compare rerank scores against the original ordering; when not enabled, record that this path did not execute |
| Citation carry-through check | Compare each answer sentence against the cited source text | Using the sample source text as the baseline, check whether the draft sentence stays within the source subject, object and period |
Three common errors
- Treating a similarity score as a store or permission filter. A retrieval score reflects relevance only, and a citation for store A-012 does not automatically exclude materials from other stores because of a high score. Check: submit the same question and observe whether the returned candidates include entries from stores other than A-012; their presence indicates that the score performed no entity isolation.
- Treating a single store-visit record as a long-term state. The sample holds one March record, and a draft sentence claiming "over the long term" exceeds the evidence. Check: compare the draft sentence field by field against the source text; the period field covers March only, so the determination is unsupported.
- Treating a rectification notice pending review as completed rectification. A rectification notice in pending-review status does not demonstrate that rectification is complete. Check: inspect the notice status field; when it reads pending review, record the evidence gap and do not write it as a closed loop.
How to verify the setup
- Enter the sample source text and the draft sentence, and compare whether the subject, object and period of the draft sentence all fall within the source text. Expected: when the period field exceeds March, the determination is unsupported, and the gap is retained in the record.
- Enter an explicit paper input of an empty array [], and check whether the downstream carry-through design continues along the empty array. Expected: the path does not break because of the empty array, and no claim is made that a particular rule necessarily returns an empty result.
- Enter a candidate set containing materials from stores other than A-012, and compare the returned entries against the target store scope. Expected: record the entries actually observed; if out-of-scope entries appear, retrieval performed no entity isolation and a separate permission implementation is required.