What the documents look like
Price and display auditing happens during day-to-day operations, after opening support and franchise training have ended. The work concentrates on whether standards stay in force across ordinary trading weeks. Supervisory visit systems, POS price tables, store camera records and the head office promotion calendar form the shared basis for comparison. Supervisory visit photographs, price tag lists, display standards and rectification notices are the documents handled routinely. Operating reports and exception reviews depend on tracing a stated conclusion back to a specific piece of material. Checking price tags, capturing photographic evidence and following a rectification loop to closure consume the most time. Sources are numerous and versions change often, so citation tracing becomes the precondition for a review that can hold up.
The knowledge base search node connects one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. When nothing is retrieved, the output path still runs through. The citation limit controls, in tokens, how much content a single search may cite. The minimum relevance value is a 0-1 filter threshold that applies only under semantic retrieval or when result reranking is used; it does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled through the setting. Image retrieval requires image input and matching model capability. Retrieval on its own does not constitute authorization, and closing a rectification loop or issuing a report still requires downstream business action.
Constraints at the citation tracing step
Three operational facts shape this step. First, the same conclusion may draw on a supervisory photograph, a POS price table and a rectification notice at once, and these documents carry different subjects, dates and status values. Second, a rectification notice can record that a store manager signed for receipt while the rectification status remains pending recheck; signature and completed recheck are separate facts. Third, a retrieval result is an array of citations, and an empty array means only that no content was found. It says nothing about whether a real-world rectification finished.
A hypothetical paper exercise sample makes this concrete. A supervisor in the operations department of a chain brand submits a draft store price tag exception review in the second week of March 2026. The material available for checking reads as follows. "The supervisory visit photograph of 10 March shows the shelf price tag in zone A marked at 19.9 yuan." "The POS price table for the same day lists this item at 22.9 yuan." "The rectification notice states that the store manager signed for receipt on 11 March, and the rectification status is pending recheck." The draft sentence awaiting judgment reads: "The price tag issue at this store was fully rectified on 11 March."
The tracing task is to compare that draft sentence against the source wording. The rectification notice records a signature date and a status of pending recheck. It does not record a completed recheck or a passed result. The draft sentence therefore is not supported by the material, and the correct handling is to mark it as pending verification rather than treat it as a completed rectification. The price comparison itself is a separate check: the photograph and the POS table give two mutually exclusive values, 19.9 yuan and 22.9 yuan, for the same item, the same date and the same price tag field. That contradiction is a usable evidence item for an exception review. Neither check depends on the other.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Select the supervisory photograph base and the price tag list base | Both use the same vector model; confirm the identifier first, then link them together |
| Search mode | Hybrid retrieval | Price tag numbers and item names are precise lookups; semantic recall supplements natural language descriptions |
| Citation limit | Set to a token amount covering one store's single-week material | Compare against the actual total length of this week's material; raise it if insufficient |
| Minimum relevance | Set 0.5 as a trial starting point | Compare against the score distribution after reranking is enabled, and observe how many results remain after filtering |
| Result reranking | Enabled | When candidates are numerous, rerank scores provide a second ordering and a filter basis |
| Question optimization | Enabled | Review follow-up questions often contain references; complete them before retrieval |
The linked knowledge bases must share the same vector model. Bases using different vector models cannot be linked together; the check is to confirm the identifier and exclude those that do not match, then record what was actually observed. The minimum relevance value filters low-relevance results and does not exclude unrelated stores or unauthorized records. The citation limit is expressed in tokens, which is a different unit from the thinking budget used in AI configuration; the thinking budget is a reasoning budget and does not serve as a citation token limit.
Three common errors
- Treating retrieval output as an authorized conclusion. The search node returns a citation array, and a length of zero is permitted. An empty array means only that no content was retrieved, and it has no bearing on whether a rectification was actually completed. The check is to compare the draft sentence against the rectification status field in the rectification notice. Pending recheck and completed are different values, and the comparison input, the comparison target and the observable result must all be stated.
- Writing a pending-review status as a lack of operating permission. The rectification notice records that the store manager signed for receipt and that the status is pending recheck. Signature and completed recheck are distinct facts, and a pending review does not demonstrate that permission is absent. The check is to compare the source status wording with the draft wording character by character. When the material does not state a recheck result, the record stays as pending verification.
- Using the minimum relevance threshold as a substitute for entity scoping. The threshold filters low-relevance results only; it does not exclude unrelated stores or unauthorized records. The check is to verify whether the material to be retrieved belongs to the target store and the target week. When it does not, record the gap rather than assume the threshold handled it.
How to verify the setup
- Take the source wording above together with the draft sentence. Input the draft sentence "The price tag issue at this store was fully rectified on 11 March" and compare the rectification notice wording "pending recheck" against the draft wording "fully rectified." If the source does not state that a recheck passed, the draft sentence is not supported and should be changed to pending recheck. The expected observation is a wording mismatch on the same subject, the same period and the same field.
- Take the supervisory visit photograph wording and the POS price table wording. Input "zone A shelf price tag, 10 March" and compare the two listed prices, 19.9 yuan and 22.9 yuan. Two mutually exclusive values for the same item, the same date and the same price tag field constitute a price inconsistency, which can serve as an evidence item in the exception review. The expected observation is a same-field contradiction, not a mere difference in emphasis.
- Take a paper input and record the search node citation output as an empty array. Check whether the downstream answer step still runs through and clearly indicates that no content was retrieved. Carrying an empty result is existing node behavior; it does not mean that a fictional rule search necessarily returns an empty array. If the downstream step does not handle an empty array, record that as a design gap awaiting completion. The expected observation is whether the downstream path continues and states the absence of retrieved content.
These three checks use the sample material given above and do not claim that any test, deployment or approval has taken place. Where the material does not state a recheck result, the evidence gap remains open. Retrieval is not authorization, and the rectification loop and report issuance still require their stated downstream implementation.