What the documents look like
Franchise supervision differs from direct-operated fresh-food ordering because the head office monitors how franchisees carry out standards. The main data sources are the supervisor app, POS sales, the franchise system and store photographs. These materials update weekly and monthly. The documents handled day to day are store visit forms, franchise contracts, rectification notices and display standards.
Citation tracing addresses one direct question for operating reports and exception reviews: can each conclusion in a report and each judgment in an exception review point to a checkable source document, and does that document's version, period and authorization status support the use being made of it? A store visit form may record a finding from one inspection date. A franchise contract may set a submission deadline. A rectification notice may record that a notice was issued. These are separate facts, and the review must keep them separate.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array whose length can be zero. Even when nothing is found, the output path still runs. Empty citations and workflow failure are two different events, so an exception review cannot treat an empty array as a retrieval error.
A knowledge base search can select one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Verification first checks the identifier and excludes those that do not match; it does not run cross-model retrieval to compare missed recalls. Hybrid retrieval fuses multiple result paths through RRF as part of the retrieval process. There is no separate on/off switch for RRF. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation. Retrieval itself is not authorization.
Hypothetical exercise sample: A franchise contract scan states "Party A: a convenience store head office; Party B: franchisee Wang; contract period: 1 January 2024 to 31 December 2026; clause 7.3: Party B shall submit photographs of completed display rectification for the previous month by the 5th of each month." A March 2025 store visit form records "inspection on 12 March, refrigerated cabinet display did not meet standard, rectification notice A-0312 issued." A draft sentence awaiting judgment reads "Wang's store completed all display rectification for March 2025." The subject is Wang's store, the object is March 2025 display rectification, and the claimed status is "all completed." The available materials show that a rectification notice was issued and that the contract requires photographs by the 5th of each month, but the sample provides no completed rectification photographs or acceptance record. The draft sentence is unsupported and requires completed photographs or an acceptance record before it can be judged.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Correspondence with the sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases with the same vector model | The store visit form base and the contract base must share a vector model to be associated together |
| Search mode | Semantic retrieval, full-text retrieval or hybrid retrieval | Contract clause numbers suit full-text retrieval; rectification descriptions suit semantic retrieval |
| Minimum relevance | A 0-1 value, effective only in semantic retrieval or with result reranking | Hypothetical starting value 0.5 because the sample question has a clear subject; compare with 0.3 to observe recall range |
| Citation limit | Each search cites at most n tokens of content | Hypothetical starting value 2000 because store visit form and contract passages are short; compare with 1000 to observe truncation |
| Result reranking | Requires a ReRank model and must be enabled in settings | May be enabled when the sample has many candidate results; observe the rerank score |
| Cited content output | Array format, length can be zero | Use a paper input of [] to check downstream handling design |
Three common errors
- Treating an empty citation array as retrieval failure or workflow termination. Input a paper-constructed []. Compare whether the downstream node still executes. Observe whether the output path runs through. The documentation states the length can be zero and the path runs through.
- Treating minimum relevance as a permission or entity filter. Input the sample's store visit form for Wang's store. Compare whether retrieval results exclude other franchise store records only because of a low relevance score. Observe whether unrelated store content appears. Relevance is a filter threshold and does not isolate entities.
- Treating a pending review or an issued notice as completed rectification. Input rectification notice A-0312 from the sample. Compare whether completed photographs or an acceptance record exist. Observe whether the draft sentence "all completed" has corresponding evidence. Issuing a notice and completing rectification are two facts.
How to verify the setup
- Input clause 7.3 of the contract in the sample and the draft sentence "Wang's store completed all display rectification for March 2025." Compare the contract requirement (submit photographs by the 5th of each month) with the draft sentence status (all completed). Observe whether the materials contain completed photographs or an acceptance record for March 2025. If absent, the draft sentence remains unsupported.
- Input the store visit form record "inspection on 12 March, refrigerated cabinet display did not meet standard, rectification notice A-0312 issued." Compare the notice number with the period referenced by the draft sentence. Observe whether a closed-loop record exists for the same number. If only the notice issuance record exists, rectification completion is not judged.
- Input a paper-constructed empty citation array []. Compare whether the downstream answer node is still triggered. Observe whether the output path runs through without error. This check addresses empty-result handling design and does not claim that a fictional rule retrieval necessarily returns [].