What the documents look like
Franchisee training and opening coaching is a transfer of operating capability. It differs from franchise filing administration, which is a legal disclosure process. The working materials for this page are training sign-in sheets, opening checklists, SOP manuals, and assessment results. Retrieval here serves contract and transaction material comparison. These materials update per store or per month, and a single document may carry a subject name, store number, training date, assessment conclusion, and material list in the same record. The retrieval step needs to pull clauses or records from the SOP manual that correspond to the sign-in sheet, the checklist, and the assessment results, so that comparison can proceed.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be linked together. Check the model identifier first, exclude those that do not match, and record what is observed. Do not run a cross-model search to compare missed recalls. For search mode, compare semantic, full-text, and hybrid retrieval against the same set of test questions. Do not assume one mode suits all materials. The reference cap uses tokens as its documented unit; test with materials of different lengths to see whether content stays complete, then adjust the cap. Minimum relevance takes effect only under semantic retrieval or when result reranking is used; label relevant and irrelevant samples before tuning the threshold. Question optimization handles coreference resolution and question expansion; compare the original question with the optimized question and check whether the subject was altered. Retrieval and recall do not perform permission or entity isolation, and a similarity score is not a basis for authorization.
Hypothetical paper exercise sample: In a franchisee training file, the SOP manual states: "After cashier training is completed, the regional supervisor signs the opening checklist to confirm, and only then may independent work be arranged." The training sign-in sheet records: "March 10, 2025, cashier training, trainee Zhang, signed in." The opening checklist records: "March 12, 2025, cashier position, supervisor Li signed to confirm." The draft sentence to be judged is: "Zhang has completed cashier training and received authorization for independent work." The retrieval target is the SOP sentence about supervisor signature before independent work. Comparing the draft sentence with the source: the source requires supervisor signature confirmation, and the checklist holds a signature record, but the sign-in sheet records attendance only and states no assessment pass or authorization effective date. The draft sentence is unsupported, and the evidence gap must remain.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the SOP manual base that uses the same vector model | Verify the vector model identifier matches before linking; exclude bases that do not match and do not search them together |
| Search mode | Hybrid retrieval | Run semantic, full-text, and hybrid retrieval on the same test questions; compare whether recalled content covers the sign-in sheet, checklist, and assessment result fields |
| Reference cap | Hypothetical starting point of 800 tokens for this exercise | Test with materials of different lengths and compare against shorter or longer caps to observe whether key fields are truncated |
| Minimum relevance | Hypothetical starting point of 0.3 for this exercise | Label relevant and irrelevant samples first, then tune the threshold; this value applies only under semantic retrieval or result reranking |
| Question optimization | Enabled | For follow-up turns, compare the original question with the optimized question and check whether the subject or store number was altered |
| Result reranking | Enabled according to settings | Reranking requires a model and is enabled according to settings; compare candidate order before and after reranking to observe whether text results move forward when the question is clear |
Three common errors
- Treating a similarity score as a permission or entity filter. Input a query that contains another franchisee's store number, then compare whether the returned content includes records for that number. If records outside the current authorized scope are returned, retrieval has not performed isolation, and access control must be handled elsewhere in the business layer.
- Describing the RRF fusion of hybrid retrieval as a toggle switch. Input the same query, compare the semantic recall and full-text recall results, and observe whether RRF merges and orders the multiple result paths. Describing RRF as an independently switchable option does not match the documented description.
- Treating a pending review or pending signature as completed authorization. Input the draft sentence "Zhang has completed cashier training and received authorization for independent work," and compare the SOP source requirement against the actual sign-in sheet and checklist records. The sign-in sheet records attendance only, and the checklist holds a supervisor signature but states no assessment pass. The draft sentence is unsupported and cannot be judged as authorization taking effect.
How to verify the setup
- Input the SOP source sentence and the draft sentence from the hypothetical sample, then compare the source requirement of "supervisor signature confirmation" against the checklist signature record and the sign-in sheet attendance record. If the returned reference contains the SOP source and the draft sentence is marked unsupported, the condition is met. If the returned content treats attendance as equivalent to authorization, the condition is not met.
- Input a query that contains only a store number, then compare whether the returned content includes records for that number. If the returned content is an empty array, the query has no match in the current knowledge base. If records for another store are returned, check the linked knowledge base scope and access control, and record what is observed.
- Input a follow-up question, then compare the original question with the optimized retrieval terms and check whether the subject name, store number, and date were altered. If the optimized question keeps the original subject and adds the referenced object, the condition is met. If the subject is replaced with another franchisee, the condition is not met.
The verification above uses a hypothetical sample and does not claim that testing has been performed. Empty result output is returned as an array, its length can be zero, and the output path still runs. Image retrieval requires corresponding image input and model capability. Reranking requires a model and is enabled according to settings. The optional paths shown in the overall document flow diagram do not mean that every request executes all of them.