What the documents look like
This category covers disposable plastic bag and single-use item reporting and environmental supervision. It sits apart from cashier membership promotion work, so the material set is different. The main sources named in the brief are the Ministry of Commerce reporting system, POS plastic bag sales, the procurement system, and store ledgers. The documents handled day to day are plastic bag procurement ledgers, sales vouchers, half-year reports, and self-discipline commitments. Drafting copy for this category means checking proposed sentences against those materials: a claimed purchase quantity, sales destination, or recovery status needs a corresponding entry in a ledger or voucher.
Two update rhythms appear. The half-year rhythm belongs to the half-year report; the daily rhythm belongs to sales vouchers. Retrieval has to keep the period scope straight, because a daily voucher value is not a half-year report conclusion. The brief lists the three most time-consuming steps as counting usage, reconciling sales vouchers, and reporting recovery information. Those steps define what a copy draft must be able to trace.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases on different vector models cannot be associated together. The check is to compare vector model identifiers first and exclude any that do not match; a cross-model search is not run to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval combines multiple result paths through RRF, and RRF is a fusion process rather than a switch that can be turned on. The citation limit controls how much content a single search cites, measured in tokens. Minimum relevance is a 0-1 filter threshold that takes effect only under semantic retrieval or when result reranking is used; it filters low-relevance results and does not perform permission or entity isolation. Question optimization handles coreference removal and question expansion and adds an extra model call. Image retrieval requires image input and matching model capability, and reranking requires a rerank model enabled according to settings. These optional paths do not mean every request executes all of them.
Hypothetical paper test case. A store ledger reads: "From 1 March 2025 to 30 June 2025, this store procured 12,000 disposable plastic bags, of which 8,000 were used in the fresh produce area and 4,000 in the checkout area." A sales voucher reads: "From 1 April 2025 to 30 June 2025, 3,600 plastic bags were sold in the checkout area." The draft sentence to be judged is: "This store's first-half plastic bag procurement was entirely used for fresh produce sales." The comparison must check whether the subject, period, and field of the draft sentence match the ledger text. The ledger records a total of 12,000 bags split across two uses; the sales voucher covers only April to June and only the checkout area at 3,600 bags. The claim that everything went to fresh produce sales is not supported by the available material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate the two bases holding the procurement ledger and the sales voucher | Compare vector model identifiers first; exclude any base that does not match, and do not run a cross-model search |
| Search mode | Semantic retrieval together with full-text retrieval | Ledger dates and quantities are exact fields, checked with full-text retrieval; use descriptions are checked with semantic retrieval |
| Citation limit | Set to 2,000 tokens as the starting point for this trial | This is a hypothetical trial starting value; compare whether ledger and voucher text is cited in full, then adjust |
| Minimum relevance | Set to 0.3 as the starting point for this trial | This is a hypothetical trial starting value; label relevant and irrelevant samples first, then compare filtering behaviour |
| Question optimization | Enabled for this trial | Compare the original question with the optimized question and check whether subject and period were altered incorrectly |
| Result reranking | Enabled for this trial | Enable when a rerank model is available; compare rerank scores with whether cited content corresponds to the draft sentence |
Three common errors
- Treating "entirely used for fresh produce sales" as a verified conclusion. The ledger text records 12,000 bags procured with two uses, 8,000 for fresh produce and 4,000 for checkout. The draft sentence does not match the ledger field values, so it is unsupported and the evidence gap must be recorded.
- Treating minimum relevance as a permission or entity isolation control. This value only filters low-relevance results. It does not exclude unrelated business lines or unauthorized records, and retrieval by itself is not authorization.
- Writing RRF as a switch that can be turned on, or treating image retrieval and reranking as guaranteed on every request. These are optional paths listed in the documentation; they require the corresponding input and model capability and are enabled according to settings.
How to verify the setup
- Input the ledger text from the hypothetical sample together with the draft sentence "This store's first-half plastic bag procurement was entirely used for fresh produce sales." Compare the use field of the draft sentence with the two use values in the ledger. The expected observation is that the output reports the claim as unsupported and preserves the evidence gap, without presenting a verified conclusion.
- Input the sales voucher text "From 1 April 2025 to 30 June 2025, 3,600 plastic bags were sold in the checkout area." Compare the voucher period with the "first half" wording in the draft sentence. The expected observation is that the output points out the period mismatch and does not treat April-to-June data as the full first-half total.
- Use a paper input of an empty array to check whether the downstream chain runs when the cited content length is 0. Compare the node output with the downstream design and observe whether the empty result is handled. Do not claim that a fabricated rule makes retrieval always return an empty array.