What the documents look like
This category covers disposable plastic bags and single-use items reporting for supermarkets and convenience stores. The work sits inside environmental regulation and one-off plastics reporting, and it draws on the Ministry of Commerce reporting system, POS plastic bag sales, procurement systems and store ledgers. The documents handled day to day are plastic bag procurement ledgers, sales vouchers, half-year reports and self-discipline commitments.
For copy drafting, these materials supply the checkable facts that a consumer-facing sentence may rely on. The subject, object, status and validity scope recorded in a document define what a draft may reference. A statement that reaches past that boundary becomes unsupported content during later review. Half-year reports and daily sales data update on different rhythms, so a draft has to make clear which type of material it cites and avoid presenting a periodic figure as an ongoing condition.
Constraints at the multi-turn prompts step
In multi-turn dialogue, history carries earlier content forward, so the prompt needs to state the material scope, the answering task and how to handle missing evidence. The model should not invent business facts. Where a subject, quantity or status is not recorded in the material, it stays as an item to verify.
When a conversation continues, the system takes the session from the database by chatId, uses the last element of the messages array as the user question, and ignores the remaining messages. When a new session passes no chatId, the context is built entirely from the supplied messages. Question optimisation involves reference resolution and question expansion, and it does not apply to every short question; the context actually carried in needs checking. The model's supported context length and reply cap affect how long documents and follow-up combinations are handled. Reply format follows the downstream consumer, and choosing json_schema requires filling in the corresponding format requirement and verifying the returned structure.
Hypothetical paper exercise: a supermarket store ledger records "March 2025 procurement of 12,000 disposable plastic bags", the sales voucher shows 8,500 units sold through POS in the same period, and a half-year report draft proposes the sentence "plastic bag use at this store has continued to decline". The draft sentence under review is "plastic bag use at this store has continued to decline". Review compares the ledger procurement quantity, the POS sales quantity and the scope of the half-year report draft, and decides whether "continued to decline" is supported by clearly mutually exclusive values for the same field in the same period. This sample is a paper exercise and involves no real customer and no actual submission.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| Memory turns | Keep a small number of history turns | Customer service and knowledge-base question answering applications usually keep a small number of history turns; compare whether the context actually carried into a continued session covers the material needed for the current review |
| Reply cap | Turn on and raise moderately | Raise moderately when generating plans or longer explanations; compare whether the reply is cut off early |
| Temperature | Low | Copy drafting needs stable citation of material; compare whether different turns describe the same material consistently |
| Reply format | Default | Ordinary question answering keeps the default; change it only when a downstream flow needs to read a fixed format |
| Multimodal recognition | Enable image recognition according to model capability | Used when material contains screenshots or table images; compare whether the request is filtered by model capability before sending |
| Hide AI output | Off | Observe intermediate results during debugging; compare whether the final display matches the internal processing |
The values above are assumed starting points for this exercise, not official defaults.
Three common errors
- Writing "continued to decline" into copy as a confirmed fact. The input is the ledger procurement quantity and the POS sales quantity; compare whether the same field in the same period has clearly mutually exclusive values. If the material shows only a single period, the statement is unsupported.
- Treating earlier content in the history as evidence from the current material. The input is the context carried into a continued session; compare whether it comes from the current review scope. If it comes from another session or unrelated material, it cannot serve as a basis for the current copy.
- Treating a similarity score as a permission or entity filter. The input is the relevance value returned by retrieval; compare whether it is used to exclude unrelated records. Retrieval itself does not carry out authorisation decisions.
How to verify the setup
- Input the ledger text, the sales voucher text and the draft sentence from the hypothetical sample; compare whether "continued to decline" in the draft sentence is supported by clearly mutually exclusive values for the same field in the same period. If the material shows only single-period data, mark the item as pending verification.
- Input the chatId and messages array of a continued session; compare whether the number of history turns actually carried in matches the configured memory turns, and observe whether the last message is used as the user question while the rest are ignored.
- Input a paper scenario with an empty reference result []; compare whether the downstream node handles it as content without references. Do not claim that a fictional rule retrieval necessarily returns an empty result; if a non-empty result is actually returned, record the observed content and check the source identifier.