What the documents look like
Franchisee training and opening coaching transfer operating capability. The materials are training sign-in sheets, opening checklists, SOP manuals and assessment results. They come from a training system, a supervisor app, store photographs and a headquarters SOP library. In a cross-border, multilingual setting, one SOP manual can exist in several language versions, and one opening checklist can be completed separately by a regional supervisor and a franchisee owner. The citation tracing step answers a narrow question: can a cited statement be returned to a specific material, a specific version and a specific completing party.
Constraints at the citation tracing step
The knowledge base search node outputs citations as an array, and the length can be zero. Even when no content is found, this output path still runs through. A downstream answer step therefore has to handle an empty citation array on its own; an empty array is not by itself a workflow failure.
The search node can associate one or more knowledge bases only when they share the same vector model. Check the knowledge base identifier and the vector model identifier first and exclude those that do not match. Retrieval is not authorization. A similarity score is a relevance value, not a category, entity or permission filter, and it does not exclude unrelated businesses or unauthorized records. Whether a citation can be used still depends on review against the effective version.
Hypothetical paper exercise sample. The English SOP manual, section 3.2, reads: "Opening checklist must be signed by the regional supervisor before the store starts operation." The Chinese SOP manual, section 3.2, reads: "The opening checklist must be signed by the regional supervisor before the store may start operation." (English rendering) A June 2025 training sign-in sheet records that franchisee owner A attended cashier training and that regional supervisor B signed on 10 June 2025. A draft sentence under review reads: "The store's opening checklist has been signed by the regional supervisor, and the store may start operation." (English rendering) The task is to decide whether this draft sentence is supported by the materials.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the English SOP library and the Chinese SOP library, after checking that both carry the same vector model identifier | The documented condition is that only knowledge bases with the same vector model can be associated together; if the identifiers differ, exclude them and do not run a cross-model search to compare missed recall |
| Search mode | Hybrid search | The sample contains a numbered section and proper nouns, so full-text retrieval and semantic retrieval are used together and merged through RRF |
| Minimum relevance | Assumed starting value 0.5 | This value takes effect only under semantic retrieval or result re-ranking; compare against 0.3 and 0.7 and observe changes in citation count and content |
| Citation limit | Assumed starting value 2000 tokens | The documented control is expressed in tokens; compare against 1000 and 4000 and observe whether citations are truncated |
| Result re-ranking | Enabled, after checking that a re-rank model is configured | Re-ranking produces a 0-1 score that can be used for filtering; when no model is configured, it is not enabled |
| Question optimization | Off | The sample question already has a complete subject and predicate, so turning it off avoids an extra model call and added latency |
Three common errors
- Treating an empty citation array as a workflow error. The documented output length can be zero, and the output path still runs through. Check by feeding a clearly labelled paper input of an empty array and observing whether the downstream step still accepts it under an empty-citation design. Do not claim that a hypothetical rule search necessarily returns an empty array.
- Treating a similarity score as a permission or entity filter. The minimum relevance value filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. Check by running the same question and observing score changes together with citation content, confirming that the score carries no authorization decision.
- Treating retrieval output as an already-signed fact. In the sample, the sign-in sheet records that regional supervisor B signed on 10 June 2025, while the draft sentence states that the opening checklist has been signed by the regional supervisor. Compare the subject, object and recorded fields of the draft sentence with the sign-in sheet. The sheet does not state the signing status of the opening checklist, so the draft sentence is unsupported.
How to verify the setup
- Input the English SOP section 3.2 text and the Chinese SOP section 3.2 text, then compare whether both versions are hit in the citation output. Observe whether the corresponding material identifier and version information appear in the citation array. If they do not appear, record the gap.
- Input the draft sentence "The store's opening checklist has been signed by the regional supervisor, and the store may start operation." (English rendering) and compare it with the fields recorded on the sign-in sheet. The sheet records only cashier training and a signing date; it does not state the signing status of the opening checklist. The draft sentence is judged unsupported.
- Input a clearly labelled paper empty array and observe whether the downstream answer step still accepts it. If the downstream step treats the empty array as an error, record that design difference. Do not claim that an actual search has returned an empty array.