What the documents look like
In franchise supervision, the head office monitors how franchisees carry out agreed standards. Order fulfilment and after-sales answers draw on four document types at once: store visit forms, the franchise contract, rectification notices, and display standards. These materials arrive from the supervision app, POS sales, the franchise system, and store photographs, and they refresh weekly or monthly. A single order dispute can therefore involve store photographs and visit records from different weeks. Because the consequences include franchise disputes and contract breach, any cited content must trace back to a specific document version and a specific clause. Every sentence about fulfilment status or after-sales responsibility should have matching source text in those materials.
Constraints at the citation tracing step
The knowledge base search node returns citations as an array, and the length may be zero. This means the output path still completes even when nothing is found. When checking an answer, separate "retrieval returned nothing" from "the workflow stopped with an error." Only knowledge bases that share the same vector model can be linked together. Before linking, confirm each base's vector model identifier and exclude any that do not match. Do not run cross-model retrieval to compare missed recalls.
The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It filters low-relevance results. It does not act as a permission filter or an entity separator, and retrieval by itself is not authorization. The document overview flow lists image retrieval, question optimization, and reranking as optional stages. Not every request executes all of them. Image retrieval requires image input and the corresponding model capability; reranking requires a rerank model and must be enabled in settings.
Hypothetical paper exercise: A franchise contract annex states, "Party A shall complete the display adjustment and return photographs within seven working days of receiving the store rectification notice." A rectification notice dated 10 March 2025 states, "Complete the end-cap display adjustment by 20 March 2025." A store photograph uploaded to the supervision app on 18 March 2025 shows the end cap still in the old display. A draft reply sentence reads, "This store completed the display adjustment within the rectification period, and fulfilment is normal." The request under review is, "Confirm that this reply can be sent directly to the franchisee."
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Store visit form base, franchise contract base, rectification notice base, display standard base | All four must share the same vector model; confirm identifiers before linking together |
| Search mode | Hybrid search | Contract clauses contain numbers and proper nouns; semantic and full-text paths merge through RRF |
| Minimum relevance | 0.4 | Hypothetical starting point to filter low-relevance fragments; compare against recall counts with the value disabled |
| Citation limit | 2000 tokens | Hypothetical starting point to cover contract clauses and rectification notice text; compare against truncation at 800 tokens |
| Result reranking | Enabled | Secondary ordering when candidates are numerous; compare ranking positions against the disabled state |
| Reply format | Default | The reply is a natural-language explanation and does not need a fixed json_schema structure |
Three common errors
- Treating a retrieval hit as an authorization basis. Finding a contract clause in a linked knowledge base only shows the clause exists in that base. It does not show the supervisor has permission to send an interpretation of that clause to the franchisee, and it does not show the contract is effective or the rectification is complete. Compare the authorization state behind the reply against the actual approval record. When either is missing, record it as pending verification.
- Treating an empty citation array as a system failure. The search node output length may be zero, and an empty array is not the same as a workflow error stop. Compare the node output against the downstream design and confirm whether the downstream handles an empty-citation branch. When no such branch is designed, record it as a gap.
- Treating minimum relevance as a permission or entity filter. This value only filters low-relevance results and does not exclude unrelated business records or unauthorized records. Compare the citation items returned before and after filtering, and confirm the filtered items were removed for insufficient relevance. Unauthorized records still require separate review.
How to verify the setup
- Using the rectification notice text "Complete the end-cap display adjustment by 20 March 2025" and the store photograph record "18 March 2025, end cap still in old display" from the hypothetical exercise as input, compare the draft sentence "completed the display adjustment within the rectification period" against the state recorded in those two materials. If the draft sentence and the photograph record disagree on display state for the same store, the same rectification item, and the same period, the sentence is not supported by the materials. Mark it as pending verification and preserve the evidence gap.
- Using the contract annex text "Party A shall complete the display adjustment and return photographs within seven working days of receiving the store rectification notice" from the hypothetical exercise as input, compare the obligated party named in the clause against the performing party described in the draft sentence. If the clause names Party A and the reply does not identify the party, the fulfilment attribution is not supported. If the clause period can be matched to the dates in the rectification notice, the actual time of the returned photographs still requires checking.
- Using an explicit paper input of an empty array [] as input, compare the downstream answer stage's handling design for empty citations. If the downstream still generates a reply containing source descriptions when it receives the empty array, the design does not meet the empty-citation branch requirement. If the downstream follows a no-citation branch and indicates that no content was retrieved, the handling design meets the condition. This check does not claim that a fictional rule search necessarily returns an empty array.
The public technical facts reviewed for this feature are: the linked knowledge base scope requires confirming selected bases and material versions, and only bases with the same vector model can be linked together; the search node outputs a citation array whose length may be zero, and an empty array does not mean the workflow has stopped with an error; the citation output is passed to the answer stage, and each reply sentence is checked against the cited source text; traceable samples are used to verify source display, and the thinking process or fluent wording does not substitute for source evidence; current and historical document versions are used to test sources, and whether a citation is adopted still depends on review of the effective version.