What the documents look like
Franchise supervision centers on whether franchisees carry out head-office standards. The working materials are store visit forms, franchise contracts, rectification notices, and display standards. They arrive through a supervision app, POS sales, the franchise system, and store photographs. Operating reviews and anomaly reviews pull checkable fields from these materials: whether a given store met the display standard in a given week, whether a rectification notice has been closed out, whether head-office orders and POS records agree within the same period.
The update rhythm is weekly or monthly. That rhythm sets the retrieval scope for a review: the search should cover documents inside the corresponding period, and records from different weeks should not be merged into one judgment. Franchise disputes and brand damage are the downstream consequences of getting this wrong, so a retrieved result has to trace back to a specific document and a specific field. A similarity score alone does not carry that traceability.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The identifier must be checked first and non-matching bases excluded; cross-model retrieval is not run in order to compare missed recalls.
Search modes are semantic search, full-text search, and hybrid search. Hybrid search merges multiple result paths through RRF. RRF is a fusion process, not a switch. The minimum relevance value takes effect only under semantic search or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation; retrieval itself is not authorization. The citation limit is measured in tokens. The cited content output is an array whose length can be 0.
Hypothetical paper exercise sample: a franchise contract annex states "Store number A-07, contract term 1 January 2024 to 31 December 2026, display standard edition March 2024 applies." A store visit form dated 6 May 2024 records "Store A-07, second shelf of the refrigerated cabinet does not meet the March 2024 edition standard, rectification notice issued." A rectification notice receipt dated 20 May 2024 records "Store A-07 rectification complete, photographs attached." The draft sentence to be judged is: "Store A-07 completed the May 2024 display rectification, and execution within the contract term meets requirements."
Checking that draft sentence requires the same subject, the same period, and the same field. The contract term covers May 2024, the store visit form states non-compliance, and the receipt states rectification complete. The receipt is material submitted by the franchisee. Whether a supervisor confirmed it, and whether it constitutes closure, is not stated in the materials and remains to be verified.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the store visit form, rectification notice, and display standard bases | First check whether the vector model identifiers of the three bases match; if they differ, do not associate them together |
| Search mode | Run semantic search and full-text search once each, then try hybrid search | Compare full-text hits on identifier-like terms such as "A-07" and "March 2024 edition" against semantic hits on "display non-compliance" |
| Citation limit | Hypothetical starting point of 3000 tokens | Compare whether cited content is truncated using store visit form passages of different lengths, then adjust |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant passages in the sample first, then compare whether the store visit form text survives filtering |
| Question optimization | Off at first, tested with complete questions | Compare the original question with the optimized question, and check whether "A-07" and the period were altered |
| Result reranking | Not enabled at first | When candidates are few, compare ordering before and after enabling against the sample fields |
Three common errors
- Treating a retrieval hit as authorization confirmation. The input is the draft sentence "rectification complete"; the comparison object is the receipt text, which only states "rectification complete, photographs attached" and does not state supervisor confirmation. The observable result is that retrieval can return the receipt, but the receipt alone does not constitute a closure conclusion, and the evidence gap must be recorded.
- Judging values from different periods or different fields as directly contradictory. The input is the May 2024 store visit form stating "non-compliant" and the 20 May 2024 receipt stating "rectification complete"; the comparison object is the same store and same field at different dates. The observable result is that these are successive stages, not mutually exclusive values of the same field at the same time, and they cannot be judged contradictory.
- Associating across vector models and then asserting missed recall. The input is the vector model identifiers of the three bases; the comparison object is whether the identifiers match. The observable result is that when identifiers differ, the non-matching base is excluded and recorded, cross-model retrieval is not run, and no run result is inferred.
How to verify the setup
- Input the draft sentence from the sample, "Store A-07 completed the May 2024 display rectification," and compare the store visit form text and the receipt text returned by retrieval. The observable result is that the store visit form text is recalled and contains "does not meet the March 2024 edition standard," and the receipt text is recalled and contains "rectification complete, photographs attached." If only one of the two is returned, record it as pending verification.
- Input the explicit paper empty result
[]and compare how the downstream node is designed to handle empty citations. The observable result is that the output path runs through. Do not claim that a fabricated rule search necessarily returns[]. - Input the vector model identifiers of the three bases and compare whether the identifiers match. The observable result is that matching identifiers allow association together, while differing identifiers lead to exclusion and a record of the actual observation, without inferring missed recall or redundancy.