What the documents look like
Franchise supervision for convenience stores centres on four document families: store visit forms, franchise contracts, rectification notices and display standards. These records are updated weekly or monthly. In cross-border work the same document set often exists in a headquarters language and in the target market language, so one question may need to reach material written in a different language. The parties named in these records are the headquarters, regional supervisors and franchise store owners. The main data sources are the supervision app, POS sales, the franchise system and store photographs. The retrieval step only handles content that has already been stored and trained; it does not create records, and it does not decide who may read or act on a record.
Semantic retrieval supports cross-multilingual understanding, so a question written in Chinese can match an English knowledge point. That capability describes content similarity. It does not describe permission, ownership or execution authority.
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. Search modes are semantic retrieval, full-text retrieval and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, then a ReRank model performs a second ordering, and the final output passes through similarity filtering and a reference limit before becoming cited content. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. Question optimisation performs reference resolution and question expansion and adds one model call before retrieval. Image retrieval requires image input and matching model capability; without a configured vision model or a vector model that supports images, the image path does not participate. Cited content is output as an array whose length may be zero, so the downstream chain still runs when nothing is found.
Hypothetical paper exercise. A franchise contract annex states: "Party A is the headquarters; Party B is a franchise store owner in the South China region; the contract is valid until 31 December 2026; the display standard in force is the March 2025 version." A store visit form records: "June 2025: the store's end-cap display did not conform to the March 2025 standard; a rectification notice was issued with a seven-day rectification period." Draft sentence to be assessed: "The franchise contract was terminated in June 2025." The comparison must use the same subject, the same period and the same field. The contract text states only a validity end date of 31 December 2026 and states no termination fact, so the draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | A Chinese store visit form base and an English contract base, both using the same vector model | Check the vector model identifier first; associate only when the identifiers match, and exclude those that do not |
| Search mode | Full-text retrieval for identifier-type questions; hybrid retrieval for natural-language questions | Compare the retrieved content of the three modes using the same test question |
| Reference limit | Hypothetical starting value of 2000 tokens | Test content completeness with store visit form and contract fragments of differing length, then compare against larger and smaller values |
| Minimum relevance | Hypothetical starting value of 0.5 | Label relevant and irrelevant samples first, then compare filtering results against adjacent thresholds |
| Question optimisation | Enabled for multi-turn follow-up questions; disabled for standalone clear questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reordering | ReRank enabled when candidates are numerous | Compare ordering before and after reordering; the reordering score can be used for filtering |
Three common errors
- Treating retrieval as authorisation. Entering "South China franchise store owner contract" and matching a contract fragment shows only that the content is similar. It does not show that the requesting party may view or execute that contract. The check compares the retrieval output against authorisation records; when authorisation records are absent, record the gap.
- Treating an unstated field as a negative. The contract text gives only a validity end date of 31 December 2026, while the draft sentence states termination in June 2025. This is a same-field value inconsistency awaiting verification, and it cannot be judged as a completed termination.
- Forcing an association across vector models. When two knowledge bases have different vector model identifiers, the node cannot associate them together. Check the identifiers, exclude the non-matching base and record the actual observation. Do not infer missing recall or redundant results from this.
How to verify the setup
- Enter the hypothetical contract text and the draft sentence above, and compare the value of the contract validity field. The expected observation is that the draft sentence is not supported, and the evidence gap must be retained.
- Enter the retrieval term "end-cap display does not conform to standard" and compare the retrieved content from the store visit form base and the display standard base. The expected observation is that the corresponding fragment is returned. If an empty array is returned, check whether the data has completed training and check the reference limit.
- Enter a Chinese question that exists only in the English contract base, and compare retrieval results from the Chinese base and the English base. The expected observation is that semantic retrieval can match the English knowledge point across languages. If no match occurs, check the vector model and the minimum relevance setting. Do not conclude that the question has no answer.