What the documents look like
Franchise supervision in convenience retail centres on headquarters oversight of how franchisees apply agreed standards. Four document families enter comparison work at the same time: store visit forms, franchise contracts, rectification notices, and display standards. A store visit form records what a store was observed doing in a given period. A franchise contract states the agreed clauses. A rectification notice carries a problem and a deadline. A display standard provides the reference against which observations are read. The brief lists supervision apps, POS sales, franchise systems, and store photos as data sources, with updates mainly weekly and monthly; this page does not derive batch synchronisation or real-time mechanisms from that cadence.
These materials belong to different parties, different periods, and different fields. Comparison only becomes meaningful once the same store, the same period, and the same field are locked together. Otherwise a difference between two documents cannot be classified as a contradiction or as unsupported. Retrieval itself does not decide authorisation, validity, or identity; it returns candidate content and orders it. The knowledge base search node can associate one or more knowledge bases that use the same vector model, and bases with different vector models cannot be associated together. Semantic search, full-text search, and hybrid search each suit different material; identifiers, model numbers, and proper nouns fit full-text or hybrid retrieval. The minimum relevance value takes effect only under semantic search or when result reranking is used, and it is a filter threshold, not a permission or entity isolation mechanism. The reference limit is measured in tokens. Question optimisation performs coreference resolution and question expansion and adds one model call. Referenced content is output as an array whose length can be zero, and the downstream path still runs on an empty result. Image retrieval requires corresponding image input and model capability, and reranking requires a model enabled in the settings. Optional paths shown in the overall flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: the source text to be checked reads — "Franchise contract clause 4.2: Party B shall submit the previous month's display self-check form before the 5th of each month. Rectification notice NO.2024-118: the store's March display problem is to be fully rectified before 10 April. Store visit form 2024-03-28: end-cap display meets the standard; cold cabinet temperature record missing." A draft sentence to be judged reads: "The franchise store's March display problem was fully rectified before 10 April, and the contractual obligation has been performed." The question is whether the draft sentence is supported by the material above.
Constraints at the retrieval step
The retrieval step returns candidates and orders them. It does not decide whether a rectification was completed, whether a contract obligation was performed, or whether a party was authorised to act. The knowledge base search node accepts one or more knowledge bases sharing the same vector model; bases whose vector model identifiers differ are checked and excluded rather than associated. A cross-model retrieval is not run in order to compare what was missed.
Search mode selection is a comparison decision rather than a fixed setting. Identifiers such as "NO.2024-118" depend on keyword matching, which full-text or hybrid retrieval provides, while natural-language phrasing depends more on semantic similarity. The minimum relevance value applies only under semantic search or when reranking is enabled, and it filters low-relevance results; it does not exclude unrelated business lines or unauthorised records. The reference limit controls how many tokens of content are referenced per search, chosen in tokens because chunk lengths vary widely across mixed knowledge bases. Question optimisation resolves references and expands the question, adding one model call; it can be turned off when the question already carries its subject. Reranking produces a 0–1 score and is merged with vector and full-text results through RRF, which is a fusion process rather than a switch. Image retrieval needs image input and a capable model; reranking needs a rerank model enabled in the settings.
Configuration and implementation checks
The table below uses the hypothetical sample above. The numeric starting points are hypothetical trial starting points for this exercise, with the reason and the comparison method stated; they are not official defaults.
| Configuration or check item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one base that uses the same vector model | Check each base's vector model identifier first; exclude those whose identifiers differ rather than associating them together |
| Search mode | Run once with semantic search, then once with full-text search | Compare what each mode recalls for the identifier "NO.2024-118" in the sample |
| Reference limit | Start at a token amount sufficient to cover the three source passages in the sample | Observe whether the referenced content includes subject, object, and status completely; raise it if not |
| Minimum relevance | Start at a middle value, effective only under semantic search or reranking | Mark relevant and irrelevant fragments in the sample first, then compare what is retained after filtering |
| Question optimisation | Turn off first, since the sample question already contains its subject | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enable when the sample produces many candidates | Rerank scores run 0–1; compare ordering changes by score |
Three common errors
- Treating retrieval recall as an authorisation decision. In the sample, the rectification notice states only that rectification is to be completed before 10 April, which is a requirement and a deadline; it does not state that rectification was completed. Reading the recalled notice as proof that the obligation was performed oversteps the retrieval step, which produces no authorisation or validity conclusion.
- Assembling materials from different fields or periods into a contradiction. The store visit form is dated 2024-03-28, and the rectification notice concerns a March problem; the fields differ. A contradiction requires explicitly mutually exclusive values in the same field. Where a value is unstated, or only stage-related evidence exists, the result is recorded as unsupported or pending verification.
- Forcing knowledge bases with different vector models into a shared association. The knowledge base search node requires associated bases to use the same vector model. When identifiers differ, the only action is to check and exclude; no inference may be drawn about behaviour after a shared association.
How to verify the setup
- Input the three source passages and the draft sentence, and compare the draft's claims of "fully rectified" and "obligation has been performed" against what the source states. The source gives only a rectification requirement and a deadline, and does not state completion, so the expected judgement is unsupported, with completion evidence still required.
- Input the identifier "NO.2024-118" and run semantic search and full-text search separately, comparing whether each mode recalls the notice containing that identifier. Full-text retrieval depends more on keyword matching, so the identifier is expected to be recalled; if semantic search does not recall it, the observation is recorded without inferring that the record does not exist.
- Use a paper input of an empty array to check the downstream design, confirming that the path still runs when referenced content has length zero. Empty-result handling can be checked with a clearly stated paper input; no claim is made that a fictional rule search necessarily returns an empty array.