What the documents look like
Copy drafting for supplier account settlement draws on purchase contracts, delivery acceptance notes, rebate agreements, and reconciliation statements. These materials update weekly or monthly. The payment relationship may be purchase-and-sale or consignment sale, and the supplier does not necessarily operate on site. Settlement method, rebate deduction rate, accepted quantity, and return discrepancy that appear in drafted copy must each be traceable to a field in these materials before they can be written.
Multi-turn dialogue carries fragments from earlier turns into later requests. When an earlier fragment is a superseded reconciliation statement or an unconfirmed acceptance note, the drafted settlement description may not match the current account state. The practical consequence is account dispute, supplier suspension of delivery, or a fairness-in-trade concern.
A hypothetical walkthrough sample is used throughout this page. The source text is: "Purchase contract HT-2024-031, supplier A, settlement method purchase-and-sale, rebate calculated at 2% of monthly purchase amount, valid January 2024 to December 2024." "Delivery acceptance note YS-2024-118, dated 12 March 2024, accepted quantity 480 cases, discrepancy 2 cases pending handling." The proposed draft sentence is: "Supplier A's March 2024 rebate is calculated at 2%, and the acceptance discrepancy has been fully handled." The question is whether "fully handled" has support in the materials. The acceptance note states only "discrepancy 2 cases pending handling" and records no handling outcome, so that phrase is not supported.
Constraints at the multi-turn prompts step
Memory turns control how much prior conversation the AI references. A larger value brings earlier acceptance notes or rebate agreement fragments into context, and the drafted copy may cite a deduction rate or quantity that has since changed. A smaller value causes a follow-up such as "which agreement does this rebate follow" to lose the earlier subject, and the answer cannot locate the specific contract. The prompt must state the material scope, the answering task, and how to behave when evidence is missing, so the model does not supply settlement facts that were not provided.
Reply format follows downstream consumption. When a later step reads a fixed structure, select json_schema and fill in the corresponding format requirement. A similarity score in retrieval results indicates textual closeness only; it does not act as a supplier permission filter or a contract validity filter.
The AI settings dialog shows only the options the current model supports. A model without multimodal capability does not present multimodal recognition options, and a model without thinking configuration does not present those options. Model selection shows maximum context, tool calling, and multimodal capability. The temperature setting controls answer stability; lower values suit rule-bound settlement copy, higher values suit creative writing. Topp overlaps with temperature, and the documentation advises against adjusting both at once. Stop sequences end output at a fixed marker and are not needed for ordinary chat. Reply length limit controls how long one answer may be. Memory turns, temperature, Topp, stop sequences, and reply format are all documented settings. None of these settings establishes authorization, approval, or a database write.
Configuration and implementation checks
The table lists documented settings and implementation checks relevant to drafting copy from these materials. The hypothetical starting values are for this paper comparison only and are not official defaults.
| Configuration item | Hypothetical starting value for this trial | Reason and comparison method |
|---|---|---|
| AI model | Select a model whose maximum context covers the contract and reconciliation fragments | Compare the model maximum context against the total length of this sample; a model without the needed length cannot hold both fragments |
| Memory turns | Set to 2 turns | This sample involves one contract and one acceptance note; compare against 6 turns, which brings in an older reconciliation fragment |
| Reply length limit | Enabled, set to a medium length | Copy must cover settlement method and rebate rate; compare against a low value that truncates the answer |
| Temperature | Set low | Settlement copy must restate material fields stably; compare against a higher value that rewrites quantities |
| Reply format | Keep default | No downstream structured read in this sample; if json_schema is selected, fill in the format requirement and verify the returned structure |
| Multimodal recognition | Not enabled in this sample | Sample materials are text fields; if enabled, confirm model support and that unsupported types are filtered before the request |
Three common errors
- Setting memory turns too high brings a rebate agreement fragment whose validity period has ended into context, and the drafted copy cites the old deduction rate. Input the sample contract together with a 2023 agreement, then compare whether the rebate rate in the reply comes from the 2024 contract. A reply showing the old rate indicates the history carry-over is not controlled.
- Treating a retrieval similarity score as a supplier permission filter. Input supplier A's contract, then compare whether the returned results mix in another supplier's acceptance note. A mixture indicates that similarity does not perform entity isolation; the identifier must be checked separately and non-matching records excluded.
- When the prompt does not state how to handle missing evidence, the model confirms "the discrepancy has been fully handled" directly. Input the sample acceptance note text, then compare whether the reply restates the "pending handling" status. A reply stating "handled" indicates the prompt does not constrain supplementation, and the prompt must state that missing evidence is retained as pending verification.
How to verify the setup
- Input the two source passages, contract HT-2024-031 and acceptance note YS-2024-118, together with the proposed sentence "Supplier A's March 2024 rebate is calculated at 2%, and the acceptance discrepancy has been fully handled." Compare whether the rebate rate in the reply matches the contract and whether the discrepancy status matches the acceptance note. The expected result is a matching rebate rate and a discrepancy status marked pending verification; a reply stating "handled" does not satisfy this.
- Input two consecutive follow-up turns under the same chatId. The first turn asks "which agreement does the rebate follow," and the second asks "what is the rate." Compare whether the second reply still points to HT-2024-031. The expected result is that the second turn retains the contract identifier; a reply that loses the subject or points to another agreement does not satisfy this.
- Input the sample contract together with another supplier's acceptance note, then compare whether the returned results contain the other supplier's identifier. If a mixture appears, record the actual observation, check the identifier, and exclude non-matching records. Do not claim that a similarity setting automatically excludes unrelated suppliers. For the empty-result handoff, a paper input of [] can check the downstream design; do not claim that a fictional rule search necessarily returns an empty list.