What the documents look like
Price and display audits take place during day-to-day operations. Regional supervisors visit stores on a weekly or monthly cadence, and those visits produce store photographs, price tag lists, display standards, and rectification notices. These materials enter a knowledge base, and retrieval requests usually come from internal training scenarios involving regional supervisors or store managers: a question about the price tag requirement for a particular promotion window, a question about a display standard, or a question about the current status of a rectification notice. Source materials are updated weekly or monthly, so a retrieval result reflects the version that was stored at ingestion time.
A rectification notice carries a status field, and that status can change over the life of the notice. A notice recorded as pending review is not an approval, and a retrieval hit on the notice text does not make the notice approved. Similarly, a price tag list for one store class does not describe the effective window for another store class. The retrieval layer returns stored text; it does not interpret the business state of that text.
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. Retrieval recalls candidates through semantic search, full-text search, or hybrid search; hybrid search merges results through RRF, and result reranking requires a rerank model and must be enabled according to the configuration. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used.
Relevance and similarity do not perform authorization or entity isolation. A retrieval hit on a rectification notice does not mean the store has been authorized to carry out the rectification, and it does not mean the notice has been approved. Retrieval is not authorization. The same-vector-model requirement for associated knowledge bases applies to the knowledge bases themselves; it does not extend to any external system.
Hypothetical paper exercise sample: internal training material for a chain operations department contains three source sentences. Sentence one: "The June 2024 East China promotion window requires Class A stores to replace price tags uniformly, effective 1 June to 30 June." Sentence two: "The price tag replacement window for Class B stores is 15 June to 30 June." Sentence three: "Rectification notice number ZG-2024-0612, status is pending review." A draft sentence to be checked reads: "Class A stores in East China begin the new price tag on 1 June, and this rectification notice has been approved." The first two sentences support the 1 June start for Class A stores. The third sentence shows only a pending review status. The second half of the draft sentence is not supported by the supplied text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the training material knowledge base that uses one vector model | Check the vector model identifier of each knowledge base first, exclude any that do not match, and do not run cross-model retrieval |
| Search mode | Hybrid search | Compare the candidate sets returned by semantic search, full-text search, and hybrid search for the same question |
| Citation limit | 800 tokens as a hypothetical starting point for this exercise | Test with price tag lists and display standards of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | 0.35 as a hypothetical starting point for this exercise | Label relevant and irrelevant samples first, then compare the returned set as the threshold changes |
| Question optimization | Enable first, then compare the original question with the optimized question | Check whether the subject is altered, such as whether "Class A stores" is replaced |
| Result reranking | Enable when a rerank model is available | Compare the ordering before and after enabling, and confirm the candidate set is sufficient |
Three common errors
- Treating a retrieval hit as authorization. Input "rectification notice ZG-2024-0612", and the retrieval returns the notice text. The observable result is that the returned content contains only "status is pending review". If the returned content is then used to conclude that the store has been authorized to carry out the rectification, the input and the comparison object are mismatched. A pending review does not constitute approval.
- Associating knowledge bases that use different vector models. Input two knowledge bases whose vector model identifiers differ, and compare whether the identifiers match. If they are still associated together and a conclusion is drawn about missed recalls or redundant returns, that asserts an undocumented runtime result. The correct action is to check the identifiers, exclude the ones that do not match, and record what is actually observed.
- Treating similarity as an entity filter. Input a question that concerns Class A stores, and observe whether the returned content includes Class B store fragments. Similarity does not exclude records for other stores. If the result is used to conclude that retrieval has isolated records by store, that treats relevance as an authorization filter.
How to verify the setup
- Input the three hypothetical source sentences and the draft sentence, and compare whether the second half of the draft sentence and "status is pending review" share the same subject and the same field. A contradiction is recorded only when the source text contains an explicit approval field with mutually exclusive values. Under the current evidence, the correct record is that the claim is not supported.
- Input a question containing an explicit number, and run full-text search and semantic search separately. Compare whether the returned candidates include the source text containing that number. Full-text search relies more on keyword hits. If semantic search does not return that number, record it as a semantic path miss, and do not infer that the number does not exist.
- Input a paper-level empty result [], and check whether the downstream node continues to run as designed, given that the citation content output can have a length of zero. Do not claim that a fictional rule retrieval necessarily returns []. Observe only how the empty-citation path is handled in the design.