Franchisee onboarding: citation tracing for operating reviews

Review citation tracing for operating reviews in Franchisee onboarding. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Opening coaching is a transfer of operating capability, distinct from legal filing and disclosure. The main materials are training sign-in sheets, opening checklists, SOP manuals and assessment scores, updated per store and per month. These materials sit across the training system, the supervisor app, store photographs and the headquarters SOP library. During operating reporting and exception review, a conclusion has to be traced back to a specific material and a specific version.

Sources are numerous and versions change quickly. The core work of citation tracing is checking whether retrieved source text matches the currently effective version, and whether statements in a draft answer are supported by that source text. The retrieval node outputs citations as an array whose length may be zero. An empty citation array means the search path completed without obtaining content; how the downstream handles it depends on downstream design.

Constraints at the citation tracing step

Knowledge base search can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Verify the vector model identifier first, exclude those that do not match and record the actual observation; do not run cross-model retrieval to compare missed recalls.

The citation limit controls, in tokens, the maximum content cited per search. The minimum relevance is a 0–1 filtering threshold that takes effect only under semantic retrieval or when result reranking is used. A similarity score and a relevance threshold do not perform permission or entity isolation, and retrieval results do not constitute authorization. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process and not a switch setting. Image retrieval requires corresponding image input and model capability; reranking requires a model and must be enabled by configuration. Optional paths shown in the overall document flow diagram do not mean every request executes all of them.

Hypothetical paper exercise sample: the source material reads, "In March 2025, the regional supervisor completed the opening check at Store A; item 4 of the checklist, 'checkout counter material placement,' is recorded as compliant; the signature field is supervisor Li; date 18 March 2025." A separate SOP manual, 2024 edition, page 7, states the checkout counter material placement requirement, and the 2025 edition, page 9, revises the same requirement. The draft sentence to be assessed is: "Store A's checkout counter material placement in March 2025 was completed in accordance with the latest SOP."

This draft sentence concerns subject Store A, object checkout counter material placement, period March 2025. The checklist records compliance but does not state whether it relied on the 2024 or the 2025 SOP edition. The draft sentence claims "completed in accordance with the latest SOP," and the materials provide no version correspondence. That statement is unsupported and can be judged only after version comparison material is supplied.

Configuration and implementation checks

Configuration itemHypothetical trial starting valueReason and comparison method
Associated knowledge basesSelect the training system and the SOP libraryVerify both vector model identifiers match first; compare against a non-matching base and exclude it
Search modeHybrid retrievalMaterials contain identifiers and natural-language description; compare coverage against single semantic retrieval
Citation limitHypothetical starting value 2000 tokensIntended to hold checklist and SOP passages; compare truncation against a 1000-token result
Minimum relevanceHypothetical starting value 0.3Compare against 0.5 to observe whether irrelevant fragments decrease
Result rerankingEnabledWhen candidates are numerous, compare ordering against a non-enabled result
Citation output handoffCheck whether the citation array is passed to the answering stepUse an empty array [] and a populated array to check downstream design separately

These values are hypothetical trial starting points for this exercise and are not official default values. The citation limit and minimum relevance are used with the units and effective conditions described in the documentation. Thinking budget is not used as a citation token limit.

Three common errors

  1. Treating a retrieval hit as authorization. Error symptom: the draft sentence states "completed in accordance with the latest SOP," while the actual material only records checklist compliance and does not state the SOP version. Check input: the checklist source text and the two SOP edition texts. Compare whether the draft statement's scope is covered by the source. Observable result: the draft sentence contains a version conclusion not stated in the source.
  2. Treating a pending review as already effective. Error symptom: the opening checklist signature field carries a supervisor signature, while the draft sentence states "the franchisee has confirmed." Check input: the signature field subject and the draft sentence subject. Compare whether the two are the same subject. Observable result: the draft sentence subject differs from the signing subject.
  3. Treating an empty citation array as an error. Error symptom: when search output is [], the workflow is judged to have terminated abnormally. Check input: the node output array length. Compare against the documented condition that length may be zero. Observable result: [] is an allowed output, and downstream handling requires separate checking.

How to verify the setup

  1. Input the checklist source text from the hypothetical sample and the draft sentence "Store A's checkout counter material placement in March 2025 was completed in accordance with the latest SOP." Compare whether the draft sentence contains an SOP version conclusion not stated in the source. Observable result: the statement is unsupported and requires version comparison material.
  2. Input the signature field "supervisor Li" from the hypothetical sample and the draft sentence subject. Compare whether the two are the same subject. Observable result: the signing subject is the supervisor; if the draft sentence states franchisee confirmation, the subjects differ.
  3. Input an explicit paper array []. Compare how the downstream design handles an empty citation. Observable result: [] can pass through the path; whether answering continues depends on downstream configuration. Do not claim that a fictional rule retrieval necessarily returns [].

The above is paper checking; no retrieval, deployment or approval was actually executed. Retrieval and forms do not constitute authorization. Signing, cooperation reaching agreement and contract effectiveness are distinct facts. Matters not stated remain evidence gaps.