What the documents look like
Opening support for a franchisee is a transfer of operating capability. It differs from franchise filing administration, which is a legal disclosure process. The materials handled in this category are training sign-in sheets, opening checklists, SOP manuals, and assessment scores. Each records a different thing: sign-in sheets record who attended, opening checklists record the state of materials and equipment, SOP manuals record required actions, and assessment scores record demonstrated mastery. Their field definitions are independent of one another, so a retrieval question has to land on a specific material. A useful query names the store, the date, and the item, for example whether a given day's sign-in sheet covers the cashier position, how a particular SOP manual version describes the inventory count action, or which assessment item a score corresponds to. These materials update per store and per month. That update rhythm does not change retrieval configuration; it affects version checking of the source material.
Two features of this material set shape the retrieval step. First, the four document types use different field vocabularies, so a retrieved passage rarely answers a question outside its own document type. Second, retrieval returns candidate passages, not verified business states. Attendance, equipment placement, and assessment results are recorded facts; permission activation, approval, and completion of corrective work are separate states that these documents do not carry unless a field states them.
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 using different vector models cannot be associated together. When identifiers do not match, the correct action is to check the identifier, exclude the non-matching base, and record what was observed. Running a cross-model retrieval to compare missed recalls is not a supported comparison.
The documented retrieval flow lists semantic retrieval, full-text retrieval, image retrieval, question optimization, and reranking as available paths. These are optional combinations. Image retrieval requires image input and corresponding model capability. Reranking requires a rerank model and must be enabled according to its setting. The presence of a path in the flow diagram does not mean every request executes it.
The citation output is an array and its length can be zero. Even when nothing is retrieved, the output path still runs. An empty result can be checked against downstream design using the paper input []. This does not assert that any particular query returns an empty array.
The minimum relevance score is a filter threshold. It takes effect only in semantic retrieval or when result reranking is used. It is not the final sorting rule, and it does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to business registration sites or other external systems.
Hypothetical exercise sample for checking. The source text to be compared reads: "On 10 March 2025, a franchise store opening checklist recorded: the cashier counter barcode scanner was in place; the inventory count was not completed. The same day's training sign-in sheet listed two attendees for the cashier position." A draft sentence to be judged reads: "On the store's opening day, cashier operating permissions were activated and the inventory count was completed." The draft sentence asserts two facts: permission activation and count completion. The source records equipment in place, attendees present, and the count not completed. Permission activation has no corresponding record.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only training material bases that use the same vector model | Check each base's vector model identifier first, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Test the same question under semantic retrieval and full-text retrieval separately | Compare the citation content returned under each mode; do not presume one mode suits all material |
| Citation limit | Hypothetical starting value of 2000 tokens | Test with sign-in sheets and SOP passages of differing length to see whether content is truncated, comparing against 1500 and 3000 for completeness |
| Minimum relevance | Hypothetical starting value of 0.5 | Label relevant and irrelevant samples first, then compare whether irrelevant fragments decrease at different thresholds |
| Question optimization | Enable for multi-turn follow-up questions containing references | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Enable according to its setting when a rerank model is available | Compare candidate ordering before and after enabling; the score is used for filtering and is not written as a switch |
Three common errors
- Treating a retrieval hit as an activated permission. In the hypothetical sample, the sign-in sheet shows only that staff attended. The draft sentence states that operating permissions were activated. Input: the sign-in sheet text. Comparison object: the permission record. Observable result: the source has no permission field, so the statement is unsupported.
- Writing an incomplete item as complete. In the hypothetical sample, the opening checklist records that the inventory count was not completed, while the draft sentence states it was completed. Input: the checklist text. Comparison object: the inventory count status field. Observable result: the same field holds mutually exclusive values, so the statement conflicts with the source.
- Treating a pending request as executed. In a hypothetical sample, a franchisee submits an opening support request whose status is pending review, and the draft sentence states that support has been scheduled. Input: the request status. Comparison object: the execution record. Observable result: pending review does not demonstrate execution, so the statement remains to be verified.
How to verify the setup
- Input the hypothetical sample's checklist text and the draft sentence. Compare the permission activation wording in the draft against the fields recorded in the source. When the source has no permission field, the sentence is unsupported.
- Input the hypothetical sample's inventory count status text and the draft sentence. Compare the values in the same field. When "not completed" and "completed" are mutually exclusive, the sentence conflicts with the source.
- Input the paper empty result
[]and the downstream node. Compare whether the downstream path continues with an empty citation. When the path runs through and no fabricated citation is produced, the empty-result handling design holds. All of the above are paper checks; no deployment or live test was performed.