What the documents look like
Retail checkout and loyalty operations revolve around four document families: promotional price tags, membership points rules, daily checkout reports, and exception refund forms. Checkout supervisors, CRM operations staff, and finance auditors hand these materials across shifts. Price disputes and membership benefit complaints frequently surface after the fact, which makes the authorization state of each item at handover more decisive than the retrieval result itself. Data updates occur daily or per shift. This page addresses one task only: operational authorization and handover management at the citation tracing step. It does not construct business functions that lack documentary support.
Constraints at the citation tracing step
Retrieval supplies citation content. It does not perform authorization judgment. The knowledge base search node outputs a citation array whose length may be zero. An empty array means the output path completes; it does not mean the workflow terminated with an error. The minimum relevance score is a filter threshold. It does not carry permission or entity isolation duties, and it does not exclude unrelated businesses or unauthorized records. Knowledge bases with different vector models cannot be associated together. Before association, verify the selected knowledge bases and material versions; only knowledge bases sharing the same vector model may be associated. Do not run a cross-model search to compare missed recall.
Hypothetical paper exercise sample: A store checkout supervisor writes in a handover record, "March 10, 2025, morning shift: promotional price tag A-031 executed at member price 9.9 yuan; exception refund form R-207 pending duty manager review; membership points supplementary entry request J-045 submitted but not approved." The draft handover sentence under review reads: "R-207 has been fully processed, J-045 has taken effect, execution may continue." Assessment point: R-207 is at the pending-review stage, and J-045 is at the submitted-not-approved stage. Neither has been completed.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason for selection and comparison object |
|---|---|---|
| Associated knowledge bases | Only the promotional rules base sharing one vector model | Compare identifiers against a base using another vector model; exclude the non-matching one |
| Search mode | Hybrid search | Compare numbered-query hit behavior against semantic search alone |
| Citation limit | 800 tokens | Compare citation count and truncation against 1200 tokens |
| Minimum relevance | 0.35 | Compare filtering of irrelevant fragments against 0.5 |
| Result reranking | Enabled | Compare candidate ordering against the disabled state |
| Question optimization | Enabled for multi-turn follow-ups | Compare short-question retrieval against the disabled state |
These values are hypothetical starting points for this exercise. They are not official defaults. Adjust them after comparing adjacent values.
Three common errors
- Treating a pending-review exception refund form as fully processed. Input the R-207 status field, compare "pending duty manager review" against "fully processed," and observe whether the draft handover sentence matches the source status. The source states pending review; the draft sentence is not supported.
- Treating a submitted-but-unapproved points supplementary entry request as effective. Input the J-045 approval record, compare "submitted but not approved" against "has taken effect," and observe whether an unauthorized execution statement appears. The source states submitted without approval; the draft sentence is not supported.
- Treating a retrieved promotional price tag as an authorization basis. Input the A-031 source text, compare the retrieval citation against the authorization boundary, and observe whether the citation is directly treated as executable permission. Retrieval is not authorization.
How to verify the setup
- Input the R-207 source status from the hypothetical sample, compare it against the draft sentence "has been fully processed," and observe whether the two conflict. If the source states pending review, the draft sentence is unsupported. Record the gap as awaiting verification.
- Input the J-045 approval record, compare "submitted but not approved," and observe whether the draft sentence "has taken effect" holds. The evidence gap remains an evidence gap; it is not a denial or a conflict.
- Input promotional price tag A-031 and the handover request, compare the retrieval citation scope against the authorization boundary, and observe whether retrieval output is directly used as an execution basis. A paper input of an empty citation array can be used to check downstream path design; it does not claim that a fictional rule search necessarily returns an empty array.
Additional verification items: check whether the retrieval output is passed to the answering step, and compare each answer statement against the cited source text item by item. Use traceable samples to check source display; do not substitute thinking process or answer fluency for source evidence. Test source timeliness with the current version and a historical version of the document; whether a citation is adopted still requires review against the effective version.
The documented overall retrieval flow lists optional paths. Not every request executes all of them. Image retrieval requires a corresponding image input and model capability. Reranking requires a rerank model and must be enabled according to settings. Cross-model knowledge bases cannot be associated; verify identifiers, exclude non-matching bases, and record actual observations. RRF is a fusion process for mixed results, not a toggle switch. The thinking budget is a reasoning budget and must not be repurposed as a citation token limit.