What the documents look like
A brand concession inside a shopping mall drafts product and marketing copy from four document families: the joint-operation concession contract, mall activity letters, sales receipts, and monthly reconciliation statements. The distinguishing fact of this category is that the concession accepts the mall's unified cashiering and deduction settlement, which ties promotional wording, price expressions and activity windows to the mall activity letter and the monthly reconciliation statement. Because activity letters and monthly settlement data sit on different rhythms, a multi-turn drafting session can carry an earlier activity letter's window into the current month's copy and produce wording that does not match the current reconciliation basis. The object of copy review is therefore the correspondence between the activity letter text, the contract terms, and the proposed sentence.
Constraints at the multi-turn prompts step
Multi-turn dialogue brings prior turns into the request, so the prompt needs to state the material scope, the drafting task, and how to behave when evidence is absent. When a session continues, the API takes the stored conversation by chatId and uses only the last element of the messages array as the user question; the remaining messages are ignored. A new session passes no chatId and builds context entirely from the supplied messages. The memory turn setting controls how many earlier turns the model may consult; a larger value carries more content and increases consumption. Question guide generation pulls the most recent six turns of conversation as context, and this should not be assumed to apply to every short question. A retrieval similarity score expresses textual relatedness and does not perform permission or entity isolation. The requirement that knowledge bases share a vector model is a retrieval-library condition and does not extend to external systems. Reply format follows downstream consumption; selecting json_schema requires filling in the corresponding format requirement and checking the returned structure. Thinking budget controls model reasoning strength and is unrelated to the reference token ceiling.
Hypothetical paper test case
Material text one: "This concession participates in the mall's unified cashiering from 1 May 2024 to 31 May 2024, and the deduction follows Appendix 2 of the joint-operation contract."
Material text two: "The mall activity letter states: from 18 May to 20 May, spend 300 and receive 50 off storewide; the concession must execute this in parallel, and prices return to normal after the activity ends."
Material text three: "The monthly reconciliation statement shows that on 20 May the concession's actual received amount matches the activity letter's discount basis."
Proposed sentence under review: "This concession offers 300 minus 50 storewide for the whole of May, valid long term."
The subject, object, state and effective range needed for judgment all appear in the texts above: the discount is limited to 18 May through 20 May, and neither the whole month nor long-term validity is supported.
Configuration and implementation checks
The table below lists officially documented settings or implementation checks relevant to this feature and task, tied to the sample above. The numeric values are hypothetical starting points for this paper exercise, not official defaults.
| Configuration item | Hypothetical starting value | Reason and comparison method |
|---|---|---|
| Memory turns | Hypothetically set to 4 | The sample needs the prior activity letter's window; compare with a value of 1 to see whether the 18 May to 20 May range is still reachable |
| Temperature | Hypothetically set low | Copy wording must match the activity letter; compare with a higher value for wording stability |
| Reply cap | Hypothetically enabled at a middle segment | The proposed sentence needs to be output in full; compare with a very low value to see whether it is truncated |
| Reply format | Hypothetically left at default | This task has no fixed-structure downstream consumer; compare with json_schema for returned structure differences |
| Thinking configuration | Hypothetically set to medium | The window-versus-whole-month distinction requires reasoning; compare with no-thinking to see whether range checking is covered |
| Multimodal recognition | Hypothetically image only | The sample includes an activity letter screenshot; compare with disabled to see whether the window text in the image is read |
Three common errors
- Writing the activity letter's 18 May to 20 May discount range as the whole month or as long-term valid. The check input is material text two and the proposed sentence; compare the date-range field and observe whether the proposed sentence covers a period the source does not state.
- Carrying the previous turn's activity letter window directly into the current month's copy. The check input is the history record and the current month's activity letter; compare the window fields across the two turns and observe whether a window appears that the current month's material does not support.
- Treating a retrieval similarity score as a permission or entity filter, so that a low-scoring result is assumed to be an unrelated business. The check input is the score in quoteList and the stated material scope; compare whether similarity is used to exclude unauthorized records and observe whether records are wrongly excluded.
How to verify the setup
- Input material text two and the proposed sentence, compare "18 May to 20 May" against "the whole of May" and "valid long term", and treat the setup as satisfied when the proposed sentence does not contain whole-month or long-term wording and retains the source window.
- Input material text one, material text three and a follow-up question about the deduction basis; compare whether the answer cites Appendix 2 of the contract and the reconciliation statement on the same basis, and treat the setup as satisfied when the answer does not add a deduction percentage the contract does not state.
- Input an empty-result handoff scenario, using the explicit paper input
[]to check the downstream design and observe whether downstream treats it as no reference; do not claim that a fabricated rule search necessarily returns empty. For knowledge bases on different vector models, first check the identifier and exclude non-matching ones rather than running a cross-model retrieval to compare missed recall.