What the documents look like
Franchisee training and opening coaching is a transfer of operating capability, not a legal filing disclosure. The working materials are training sign-in sheets, opening checklists, SOP manuals and assessment scores, refreshed per store and per month. They are the evidence used to judge whether a franchisee has completed training and whether a store meets opening conditions. Product information and selection questions must be answerable back to specific source text, so citation tracing sits at the centre of delivery quality: every conclusion in an answer should point back to a particular sign-in sheet, checklist or SOP passage.
Three facts shape this task. First, the material set mixes structured records (sign-in sheets, scores) with long procedural text (SOP manuals), so a single retrieval mode rarely covers both numbering queries and natural-language queries. Second, the update rhythm is per store and per month, which means a current version and a historical version of the same document can coexist in the knowledge base. Third, the evidence is operational, so a statement about training completion is not the same fact as an opening approval, and neither can be inferred from the other.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that use the same vector model, for vector search. Its output is an array of citations, and the length may be zero, which means the downstream chain still executes even when nothing was retrieved. Whether a retrieved item is actually used by the answering step must be checked item by item against the cited source text. The minimum relevance value is a filter threshold that takes effect only under semantic search or when result re-ranking is used; it does not perform permission or entity isolation. Hybrid search merges multiple retrieval paths through RRF, so the final ordering is not strictly equal to a single vector score.
Hypothetical drill sample: the source material reads, "In March 2025, franchisee Li of the East China region attended cashier and inventory training, sign-in sheet number PX-2025-0312, completed the assessment the same day, score 82." The draft sentence to be judged reads, "Li has completed all opening coaching courses and has been approved to open." The comparison must hold the same subject, the same period and the same field. The sign-in sheet covers only cashier and inventory training plus one assessment; it does not state completion of all courses, and it does not state an opening approval. The draft sentence is therefore unsupported, and the opening approval remains a pending item to be verified.
Configuration and implementation checks
| Configuration or check item | Hypothetical trial starting point for this page | Reason for the choice and comparison object |
|---|---|---|
| Linked knowledge bases | Training and SOP libraries that share one vector model | Knowledge bases with different vector models cannot be linked together; verify the identifier first and exclude non-matching ones |
| Search mode | Hybrid search | Numbering-type questions can draw on both semantic and full-text recall |
| Minimum relevance | Hypothetical starting point 0.5 | Compare against the score distribution of returned candidates before adjusting |
| Citation upper limit | Hypothetical starting point 2000 tokens | Compare against the actual cited length in answers before adjusting |
| Citation hand-off | Search output passed to the answering step | Check item by item whether the answer matches the cited source text |
| Document version check | One current version and one historical version | Determine whether a citation falls within the valid version range |
Three common errors
- Treating minimum relevance as a permission filter. Input: a cross-store query. Comparison object: the contents of the citation array. Observable result: the array may still contain material from other stores, which shows the threshold does not isolate entities.
- Treating an empty citation array as a workflow failure. Input: an explicitly written empty array. Comparison object: the downstream answer design. Observable result: the chain still executes, and empty-result handling must be designed separately.
- Substituting answer fluency for source checking. Input: a draft sentence and the source material. Comparison object: whether subject, period and field match. Observable result: items not stated in the source are judged as pending verification.
How to verify the setup
- Input the hypothetical source material and the draft sentence given above, compare the coverage of the sign-in sheet against the wording of the draft sentence, and if the stated completion of all courses and the opening approval have no corresponding field in the source, judge the sentence unsupported.
- Input a written empty array
[], compare whether the downstream node handles it according to the empty-citation design, and if the chain continues without error, this matches the documented behaviour that an empty array still passes through. - Input two knowledge bases with different vector model identifiers, compare whether the identifiers match, and if they do not match, exclude the link and do not run a cross-model retrieval comparison.