What the documents look like
Price and display auditing happens during ordinary trading periods, where the focus is sustained execution of standards. Supervision systems, POS price lists, store cameras and head-office promotion calendars together form the basis for comparison, with update frequency measured in weeks or months. Day-to-day documents include store visit photographs, price tag lists, display standards and rectification notices. The value of these materials lies in traceability: when an after-sales answer needs to state the price or display condition of a given store during a given period, the original records for that period must be located. The timeliness of the material directly determines whether a citation holds, since a historical version and the current version can produce different conclusions.
The knowledge base search node can select one or more knowledge bases that share the same vector model, and outputs citation content as an array whose length may be zero; when the array is empty the output path still runs through. Search results need to be passed to the answering step, where each answer is checked line by line against the cited original text. The minimum relevance value is a filter threshold that takes effect only in semantic search or when result reranking is used, and it does not perform permission or entity isolation. Retrieval is not authorization; forms collect values, and business verification, approval, database writes and external synchronisation each need their own downstream implementation.
Constraints at the citation tracing step
A hypothetical drill sample: a chain brand operations department receives a customer complaint on 10 June 2025 stating that on 8 June the shelf price tag at Store A showed 9.9 yuan while the till receipt showed 12.9 yuan. The supervision system store visit photograph dated 8 June shows the shelf price tag at 12.9 yuan, and the head-office promotion calendar states that the promotion period for this item ran from 1 June to 7 June. A draft sentence to be assessed reads: "Store A applied the 9.9 yuan promotion price on 8 June, and the price tag matched the system." The original text to be checked against is: the store visit photograph shows the price tag at 12.9 yuan on 8 June; the promotion calendar states the promotion period ran from 1 June to 7 June. The draft sentence and the two original texts differ in value on the same subject, the same period and the same field, so the draft sentence is unsupported.
The minimum relevance value is a filter threshold rather than a final ordering rule. After question optimisation, hybrid retrieval, image retrieval or result reranking is enabled, the final result may pass through multi-path recall fusion and does not necessarily follow a single vector similarity ordering. A relevance score is not a category, entity or permission filter, and there is no basis for treating a similarity setting as excluding unrelated businesses or unauthorised records.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model knowledge base holding store visit photographs and price tag lists | First check the vector model identifier and exclude knowledge bases with a different model; do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Price tag numbers and promotion calendar dates are exact queries; semantic and full-text paths are combined through RRF |
| Minimum relevance | Hypothetical starting value 0.5 | This value is a starting point for this drill; compare against 0.3 and 0.7 for recalled item count and proportion of unrelated fragments |
| Citation limit | Hypothetical starting value 2000 tokens | This value is a starting point for this drill; compare against 1000 and 4000 for citation completeness and truncation |
| Result reranking | Enable ReRank | When candidates are numerous, apply a second ordering; compare citation order for the same question before and after reranking |
| Question optimisation | Enable in multi-turn follow-up scenarios | Compare retrieval hits and elapsed time for the same follow-up question with the setting on and off |
Three common errors
- Treating a retrieval hit as authorised execution. In the hypothetical sample, retrieving the promotion calendar does not mean the store was authorised to sell at 9.9 yuan on 8 June. The input is the retrieval output, the comparison object is the validity period stated in the calendar against the complaint date, and the observable result is whether the two periods overlap. When the periods do not overlap, the citation cannot support the draft sentence.
- Treating minimum relevance as a permission or entity filter. The input is the relevance setting value, the comparison object is whether returned citations contain records from other stores or other periods, and the observable result is whether irrelevant records disappear after the relevance value is adjusted. This setting does not exclude unauthorised records, and a separate permission implementation is required.
- Treating an empty citation array as a process error. The input is the empty array output by the search node, the comparison object is whether the downstream answering step still executes, and the observable result is whether the path runs through. An empty array only means no content was retrieved this time; it does not mean the business has determined that no such record exists.
How to verify the setup
- Input the store visit photograph original text and the draft sentence from the hypothetical sample, compare the price stated in the draft sentence with the price shown in the photograph, and observe whether the two values differ. When they differ, the draft sentence is judged unsupported and must not be reclassified as verified.
- Input the promotion calendar original text and the complaint date from the hypothetical sample, compare the promotion period end date with the complaint date, and observe whether the complaint date falls outside the promotion period. When it does, the citation cannot evidence that the promotion price was applied on that day, and the item remains pending verification.
- Input a paper-recorded empty array explicitly marked as a search output, compare whether the downstream answering step continues to execute and provides a no-citation explanation, and observe whether the path runs through without fabricated sources. When these conditions are met, this check passes.
These checks describe expected observations only; no test, deployment, approval, refund or other business action is claimed to have occurred. Where the source material does not state an error code or an actual observation, the comparison to perform and the remaining unknown are recorded rather than a deterministic failure cause or result.