What the documents look like
Fruit-store procurement copy drafting draws on four record types: purchase orders, acceptance sheets, pesticide-residue test reports, and shrinkage records. Wholesale-market prices and supplier quotes update by day or by batch, and a grade-acceptance conclusion holds only for the batch it describes. Fruit purchasing depends on wholesale-market conditions and grade inspection, so a copy draft that names origin, grade, or test conclusion must anchor each claim to a specific batch record.
A multi-turn exchange typically starts with price and grade questions, then asks for a store poster or community-group message. When a grade conclusion or test conclusion from an earlier turn enters a later prompt, the batch scope must travel with it. Information absent from the source material stays absent from the copy.
Constraints at the multi-turn prompts step
Multi-turn dialogue carries earlier turns into later ones. The memory-rounds setting controls how many prior chat rounds the answer may reference; a small value can leave earlier information unused, and a larger value brings in more content. Copy drafting usually needs the rounds that hold grade, batch, and test conclusions, so the memory-rounds value should match the actual follow-up span.
A prompt should state the material scope, the drafting task, and how to handle missing evidence. A pesticide-residue test report covers only the submitted sample; a batch that was not submitted cannot be given a passing conclusion from that report. Reply format follows downstream needs; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure.
Hypothetical paper exercise sample: three source sentences. First, "Purchase order, 10 June 2025: Red Fuji apples, origin Shandong Qixia, grade one, quantity 200 boxes, supplier quote 85 yuan per box, quote valid until 12 June." Second, "Acceptance sheet, 11 June 2025: this batch of Red Fuji apples received 198 boxes, accepted grade one, 2 boxes returned for bruising." Third, "Pesticide-residue test report, 11 June 2025: submitted sample is this batch of Red Fuji apples, tested for common organophosphates, conclusion not detected." Draft sentence to be judged: "This store's Red Fuji apples all pass pesticide-residue testing, direct from origin, quality guaranteed." Judgment points: the test report covers only the submitted sample, so "all" exceeds the report scope; "quality guaranteed" has no corresponding field in the material; "direct from origin" is not stated in the purchase order. The sentence is unsupported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| AI model | Select a model supporting the required context length; check the displayed maximum context and tool-calling information | The three sample sentences plus follow-ups must fit within the context |
| Memory rounds | Hypothetical starting value for this exercise is 4 rounds, because grade, batch, and test conclusions sit in adjacent rounds; compare against the actual follow-up span | Below this starting value, an earlier grade conclusion may not be referenced |
| Temperature | Copy drafting may be raised; rule-checking scenarios stay at default | Judging whether a draft sentence is supported favors a lower value |
| Reply format | Ordinary copy keeps the default; when a fixed structure is needed, select json_schema and fill in the format requirement | Enable only when the judgment result needs structured output |
| Multimodal recognition | Enable only when the model supports it and the input contains images, audio, or video | The sample is text material, so no enabling is needed |
| Hide AI output | Enable when an intermediate result must pass to a later node | This page outputs copy directly, so keep it off |
The "4 rounds" above is a hypothetical starting value for this exercise, not an official default.
Three common errors
- Extending a test-report conclusion to all batches. The sample report states only that the submitted sample is this batch of Red Fuji apples, while the draft sentence says "all pass pesticide-residue testing"; the subject scope differs, so the claim is unsupported. Check: compare the submitted-object field in the report text against the subject term in the draft sentence; a scope difference means the condition is not met.
- Treating the quote validity date as a supply period. The sample says the quote is valid until 12 June, which is a quote field and not a supply commitment. Check: compare whether the draft sentence contains a supply-period statement; if the material has no corresponding field, record the gap.
- Treating the accepted quantity as the purchase-order quantity. The sample shows 200 boxes ordered, 198 received, and 2 returned, giving different values for the same field in the same batch. Check: compare the quantity fields in the two sheets against the return record and confirm which one the copy cites.
How to verify the setup
- Input the three sample sentences and the draft sentence, and compare "all," "quality guaranteed," and "direct from origin" in the draft sentence against the material fields. Expected: judged unsupported, with a note that the test report covers only the submitted sample. If the output extends the report conclusion to all batches, the condition is not met.
- Input the sample material and ask, "What are the grade and test conclusion for this batch of apples?" and compare whether the answer carries the batch and submission scope. Expected: the answer cites the acceptance-sheet grade and the report's submitted object. If the answer states the report conclusion as a whole-batch pass, the condition is not met.
- Input the paper input
[]for empty-result handling and check how the downstream design presents the no-reference case. Expected: the downstream handles it as no reference and preserves the gap note. This input is a paper check and does not claim that any retrieval actually returned an empty result.
A relevance score does not carry permission or entity-isolation duties. Retrieval libraries need the same embedding model to be associated together; check the identifier and exclude non-matching ones rather than running a cross-model retrieval to compare missed recalls.