What the documents look like
Franchise filing management works with legal disclosure and the commerce department system. It does not handle specific site engineering. The daily documents fall into four groups: franchise contracts, operation manuals, filing materials, and information disclosure documents. Updates are irregular, and some work advances on an annual cycle.
This shapes retrieval for product questions in two ways. Contracts and disclosure files carry many parties, periods, and field values. Operation manuals carry numbers and proper nouns. The recall path for these two kinds of content differs. Retrieved reference content is candidate material only. Authorization status, filing acceptance status, and contract effectiveness status each require separate verification outside retrieval.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together. When identifiers do not match, check first and exclude the mismatched base. Record only what is observed, and do not infer runtime results of a link that was not made.
Search mode should be compared per test question across semantic search, full-text search, and hybrid search. Semantic search depends on vector distance and suits natural-language phrasing. Full-text search depends on keyword hits and suits numbers, models, and proper nouns. Hybrid search uses both recall paths and merges them through RRF. RRF is a result fusion process and is not a standalone switch. Minimum relevance is a 0 to 1 filter threshold that takes effect only in semantic search or when result reranking is used. It does not perform permission or entity isolation. The reference limit is measured in tokens and controls how much content each search may cite. Question optimization adds one model call before retrieval for coreference resolution and question expansion. Compare the original question with the optimized question to check whether the subject was altered. Reranking requires a ReRank model and must be enabled by configuration. Image retrieval requires the corresponding image input and model capability. These paths do not run on every request merely because a flow diagram lists them.
Hypothetical paper exercise, not a real record: Source text one reads, "Party A is the franchisor, Party B is the franchisee, the contract term runs from March 1, 2024 to February 28, 2027, and Party B submits the store list required for filing at signing." Source text two reads, "The store list in the information disclosure document is current as of January 31, 2024, covering two directly operated stores." The draft sentence to be assessed reads, "Party B obtained the franchise authorization from March 1, 2024, and the filing materials have been accepted by the commerce department." The draft sentence asserts two facts, authorization effectiveness and filing acceptance. Neither source text states either fact.
Configuration and implementation checks
| Configuration item | Value or check | Relation to the sample |
|---|---|---|
| Linked knowledge bases | Select one or more knowledge bases using the same vector model | When contracts and disclosure files sit in different collections, confirm the vector model identifiers match |
| Search mode | Semantic, full-text, or hybrid, compared per question | Contract-period questions and store-list questions can be observed separately for recall differences |
| Reference limit | Measured in tokens; this hypothetical exercise starts at 2000 | This is a starting value for this paper exercise, chosen to observe whether both source texts are cited in full, then compared with other values before adjustment |
| Minimum relevance | A 0 to 1 filter threshold, effective only in semantic search or result reranking | Mark relevant and irrelevant samples first, then observe the threshold's effect on checking the draft sentence against the source texts |
| Question optimization | Used for coreference resolution and question expansion; can be on or off | Compare the original and optimized question to check whether subject and period were altered |
| Result reranking | Requires a ReRank model and must be enabled by configuration | When enabled, filter by rerank score and observe whether contract period and party fields are still cited correctly |
Three common errors
- Treating a retrieval hit as authorization confirmation. In the sample, the draft sentence claims Party B obtained authorization from March 1, 2024, while source text one states only the contract term start and end. It does not state the conditions for authorization effectiveness. Returning that contract fragment is a recall result and cannot be used to conclude that authorization took effect.
- Merging the disclosure document's store-list date with the contract period as one field. Source text two gives a store list current as of January 31, 2024, which is a different field from a contract term beginning March 1, 2024. The closeness of the two dates is not a contradiction, and it does not demonstrate a filing condition.
- Writing a pending review status as completed acceptance. The draft sentence uses "has been accepted by the commerce department," while the source texts contain no acceptance record, acceptance number, or acceptance date. Pending review is not completed acceptance. Missing evidence stays missing.
How to verify the setup
- Input source text one and the draft sentence, then compare "obtained authorization from March 1, 2024" in the draft against the contract term wording in source text one. When the source does not state an authorization effectiveness clause, the finding is unsupported and the evidence gap must be recorded.
- Input source text two and the draft sentence, then compare "filing materials have been accepted" against the store-list date wording in source text two. When the source does not state an acceptance fact, the finding is pending verification, not written as a negation or a conflict.
- Input an empty reference array
[]to check the downstream chain design, confirming that the output path still runs when the reference content length is 0. Do not claim that a fabricated rule will necessarily return an empty result from retrieval.
All three checks compare the draft sentence against the source texts at field level. Semantic recall and relevance do not carry permission or entity isolation judgments. Materials needed for verification must fall within an accessible target retrieval scope or an explicit manual input. When they are absent, record the gap.