What the documents look like
Retail checkout and loyalty work draws its price and entitlement information from four source systems: the POS system, the member CRM, the payment platform and the shelf-label system. Store cashier supervisors, CRM operations staff and finance auditors handle four document types each day and each shift: promotional shelf labels, member points rules, the cashier daily report and exception refund slips. These materials together determine the price and points basis that applied to a transaction at the time it occurred. Citation tracing for staff training therefore means keeping training scripts, store reply wording and handover notes tied back to the matching source material.
Promotional shelf labels and member points rules change with each promotion period. The cashier daily report and exception refund slips are produced per shift. When a training sentence cites a price or a points rule, it has to state the material version and the period of applicability. A figure that was correct in an earlier promotion period may not describe the current one, and a training sentence that omits the period cannot be checked against a specific label.
Constraints at the citation tracing step
The knowledge base search node can select one or more knowledge bases that share the same vector model. Its output is a citation array, and the length may be zero. A zero-length array does not mean the workflow has errored and stopped; the output path still runs. Retrieved content is passed to the answering step, where each answer is checked line by line against the cited source text.
Semantic search, full-text search, hybrid search, question optimization, result reranking and image retrieval are optional paths listed in the overall retrieval flow. Whether each one executes depends on the input and the settings. Image retrieval requires image input and a matching model capability. Reranking requires a model and runs only when enabled by settings. The minimum relevance value takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation limit controls, in tokens, the maximum content cited per search. RRF is the fusion step for combining multiple retrieval channels; it is not a switch that can be turned on. Thinking budget is a model reasoning-effort setting and is unrelated to the citation token limit. Knowledge bases with different vector models cannot be associated together. During checking, confirm the identifier first, exclude those that do not match, and record what is actually observed rather than inferring run results.
Hypothetical drill sample: the original text of a store's March 2025 promotion shelf label reads "This promotion period runs from 1 March 2025 to 31 March 2025. Brand A 1-litre cooking oil member price is 39.9 yuan per bottle, non-member price is 45.9 yuan per bottle, limit 2 bottles per person per promotion period." The member points rule for the same period reads "Members earn 1 point for every 1 yuan actually paid on promotional goods; corresponding points are deducted on refund." A draft sentence reads "During this promotion period, Brand A 1-litre cooking oil is settled at 39.9 yuan per bottle for members and earns points normally." The task is to decide whether this draft sentence agrees with the two source texts.
Configuration and implementation checks
Configuration here centers on checking the sample. Only settings and implementation check items stated in the official documents are listed. Where a number is needed, it is marked as a starting point for this hypothetical trial, with the comparison object stated.
| Configuration or check item | Hypothetical value or practice for this trial | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select only knowledge bases with the same vector model | Official requirement is that only the same vector model can be associated; confirm identifiers first |
| Citation limit | Set as a starting point for this trial, compared against the length of the sample material | The official document controls citation volume in tokens; the starting point needs to cover the sample source text |
| Minimum relevance | Set only when semantic search or result reranking is used | The official document states this value takes effect only in those two cases |
| Citation output handling | Pass the citation array to the answering step for line-by-line checking | Official output is an array whose length may be zero |
| Material period check | Check the current and a historical promotion shelf label separately | Materials carry a period of applicability; different versions give different conclusions |
| Question optimization | Decide whether to enable it based on whether the sample question is complete | The official document states question optimization adds one model call |
Three common errors
- Treating minimum relevance as a permission or entity isolation mechanism. Semantic search and relevance affect recall ordering and filtering only. They do not decide whether a store or a member may view a given record; permission requires downstream implementation.
- Forcing knowledge bases with different vector models to be associated together. The official requirement is that only the same vector model can be associated. When an identifier does not match, exclude that knowledge base and record it, rather than inferring missed recall or redundancy.
- Treating thinking budget as the citation token limit. Thinking budget belongs to the model reasoning-effort configuration. The citation limit is controlled separately in tokens. The two have different fields and different effects.
How to verify the setup
- Input the promotion shelf label source text, the member points rule source text and the draft sentence from the hypothetical drill sample. Compare whether the draft sentence simultaneously matches three fields: the 2-bottle limit, the 39.9 yuan member price, and 1 point per 1 yuan actually paid. When all three fields agree on the same object in the same period, the draft sentence is supported. When any field disagrees with the source text, or the source text does not state it, record the result as unsupported or pending verification.
- Input an explicit paper citation array of []. Compare whether the downstream answering step handles the empty result as designed. If the downstream step still runs the path on receiving an empty array and outputs a recognisable no-citation state, the condition is met. If the downstream step treats the empty array as an error and stops, the condition is not met.
- Input two knowledge bases whose identifiers show different vector models. Compare whether the association operation permits selecting them together. If the system prevents joint association and keeps each identifier separate, the check passes. If the interface permits selection, record the actual observation without inferring retrieval results.
This page addresses only retail checkout and loyalty / store service and internal training / citation tracing. In operational handover, which materials, operations and requests sit at which authorization state cannot be treated as retrieval or a form performing authorization.