What the documents look like
Retail plastic bag and single-use item reporting sits with disposable plastics reporting and environmental supervision, a different business line from checkout membership promotion. The main data sources are the Ministry of Commerce reporting system, POS plastic bag sales records, the procurement system and store ledgers. The documents handled day to day are plastic bag procurement ledgers, sales vouchers, semi-annual reports and self-discipline commitments. Data updates at two granularities: semi-annual and daily.
In store service and internal training, staff explain the basis of a report to customers, and training material marks the source document behind each statement. Citation tracing in this category lets every sentence about procurement volume, sales volume or recovery status point to a specific ledger or voucher. The knowledge base search node can select one or more knowledge bases that share the same vector model, and outputs a citation array whose length may be zero. Retrieval by itself is not authorization, and a form collecting values is not business approval.
Hypothetical drill sample: a store ledger records "January 2025 procurement of 3,000 plastic bags, and the sales voucher shows 2,800 sold that month." A semi-annual report draft proposes: "The store's plastic bag procurement volume and sales volume for the first half of 2025 have all been checked and are consistent." The request to be judged is "Please confirm this draft sentence can be used directly for semi-annual report submission."
In this sample the ledger covers January only, while the draft sentence claims the first half year, so the period ranges do not match. The sales voucher records sold quantity only and does not state recovery status. The phrase "all checked and consistent" lacks a recovery data source. The request is pending review, which does not mean it has been approved or that submission has been completed.
Constraints at the citation tracing step
The knowledge base search node accepts one or more knowledge bases that use the same vector model. Output is a citation array, and a length of zero means the downstream chain still runs even when nothing is retrieved. The citation content then goes to the answering step, where each answer sentence is checked against the cited original text. Search parameters document a citation upper limit measured in tokens, and a minimum relevance value between 0 and 1 that applies only under semantic search or when result re-ranking is used. Minimum relevance is a filtering threshold, not a final ordering rule; after question optimization, hybrid retrieval, image retrieval or re-ranking, results may be fused from several recall paths and do not necessarily follow a single vector similarity order.
The retrieval library must not be treated as an authorization scope. Similarity scores do not separate unrelated businesses or unauthorized records, and relevance is not an entity or permission filter. The same-vector-model requirement for joint association applies to the knowledge bases selected for this node; it does not extend to business registration sites or other external systems. For this page, the relevant configuration entries are the associated knowledge bases, the search mode, question optimization, the citation upper limit, minimum relevance and result re-ranking. Values below are hypothetical trial starting points, not official defaults.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this sample | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select the plastic bag procurement ledger base and the sales voucher base | Both use the same vector model, meeting the joint association condition. Compare identifiers against a base using a different vector model and exclude it |
| Search mode | Hybrid retrieval | The ledger contains numbers and dates; hybrid retrieval combines keyword and semantic matching. Compare number hit results against semantic retrieval alone |
| Question optimization | Enabled | In training scenarios staff questions often contain references. Compare consecutive follow-up retrieval results with the disabled state |
| Citation upper limit | Set to 3,000 tokens | Ledger and voucher passages are short; this value can hold several citations. Compare citation count against a lower value |
| Minimum relevance | Set to 0.4 | This is a hypothetical starting point used to filter low-relevance fragments. Compare recall count against 0.3 |
| Result re-ranking | Enabled | Secondary ordering when candidates are numerous. Compare ranking position against the disabled state |
These numbers are hypothetical trial starting points for this drill only, not official defaults.
Three common errors
- Treating retrieval output as an authorization basis. The knowledge base search node outputs a citation array and does not verify whether a store holds submission permission. Check: input a store ledger with no authorization and observe whether the system returns citation content only, without producing an approval state.
- Confusing the citation upper limit with a thinking budget. The citation upper limit controls how many tokens each search may cite; a thinking budget is a reasoning parameter in AI configuration. Check: open the knowledge base search parameters and the AI configuration dialog separately, and confirm the two values appear in different places.
- Treating an empty citation array as a system fault. The search node output length may be zero, and an empty array does not mean the workflow has errored. Check: use the paper input
[]and confirm whether the downstream node still runs.
How to verify the setup
- Input the ledger original text and sales voucher original text from the hypothetical sample, and compare the period range of the draft sentence against the ledger coverage. If the draft sentence claims the first half year while the ledger covers January only, judge it unsupported. The expected observation is that the reviewer marks a period mismatch.
- Input the request to be judged, "Please confirm this draft sentence can be used directly for semi-annual report submission," and compare the request status against authorization records. If the request is pending review with no approval record, judge it to be awaiting verification. The expected observation is that the reviewer records the evidence gap.
- Input the paper empty citation array
[]and compare whether the downstream answering node accepts it. If the downstream still executes, judge empty-result handling normal. The expected observation is that the workflow does not halt.
These checks describe what to compare and what remains unknown. They do not assert that any test, deployment, approval or submission has occurred.