What the documents look like
Rebate calculation happens after the sale. The working materials are the rebate policy, the sales achievement table, the reconciliation statement, and the red-letter invoice, with data drawn from ERP sales, dealer remittances, the contract ledger, and the finance system. Monthly and quarterly update rhythms shape how a multi-turn assistant behaves: the same reconciliation statement or the same period's achievement figure gets referenced across several turns, so earlier confirmed interpretations must return reliably. Prompts state the material scope, the answering task, and how to handle missing evidence, so the model does not write business facts that the source documents do not contain.
Constraints at the multi-turn prompts step
Multi-turn conversation relies on the memory turn setting to control how much history is carried in. When a session continues, the system retrieves the conversation from the database by chatId, uses the last item in the messages array as the user question, and ignores the remaining messages. When chatId is not supplied, no context is used and the supplied messages alone build the context. Prompts need to distinguish new-session input from continued-session input. Where continuous knowledge-base conversation is required, check coreference resolution and question expansion; whether a short question benefits depends on the actual material. Model and length settings need their supported context and reply configuration checked, and long-document plus follow-up combinations need observation. Reply format follows downstream consumption; selecting json_schema requires filling in the corresponding format requirement and verifying the returned structure.
A hypothetical paper exercise: a brand finance reviewer checks dealer A's February 2026 rebate for the East China region in March 2026. The reconciliation statement states "February 2026 sales achievement rate 92%, base rebate tier applies under policy clause 3." The rebate policy states "when the achievement rate is below 95%, the base rebate is calculated at 1.5%, and returned-goods amounts are deducted from achieved sales." The draft sentence to be judged is "Dealer A's February rebate has been calculated at 1.5% and can proceed directly to invoice offset." The sentence lacks a record confirming whether the returned-goods amount was deducted, so it is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Memory turns | Keep 4 turns | Customer service and knowledge-base Q&A typically retain a small number of history turns; compare whether a follow-up can still cite a confirmed interpretation at 2 turns versus 4 turns |
| Temperature | Low | A rules-based calculation scenario needs stable answers; compare against a higher setting to observe whether the interpretation drifts |
| Reply cap | Enabled and lowered | Q&A replies should be brief; compare before and after lowering to see whether truncation appears |
| Reply format | Default | Ordinary Q&A keeps the default; change to json_schema only when a downstream flow needs to read a fixed structure |
| Multimodal recognition | Decided by model capability | The option does not appear when the model does not support it; screenshot recognition requires image capability |
| Stop sequences | Not set | Ordinary chat generally does not need them; use only when output must end at a fixed marker |
Three common errors
- Treating a retrieval result or a form submission as authorized execution. Retrieval fetches material, a form collects values, and business verification, approval, database writes, and external synchronization each need their own downstream implementation. The interactive node response returns only the display configuration, without internal runtime fields such as entryNodeIds.
- Treating a pending review as a lack of operating permission, or merging signing, deal conclusion, and contract effectiveness into one fact. The phrase "already calculated" in the draft sentence requires a returned-goods deduction record; without that record it can only be marked as pending verification.
- Treating a relevance score as a permission or entity filter. A similarity setting does not exclude unrelated business or unauthorized records, and libraries across different vector models cannot be jointly associated; check identifiers first and exclude non-matching items.
How to verify the setup
- Input the reconciliation statement and policy text above, plus the draft sentence "Dealer A's February rebate has been calculated at 1.5% and can proceed directly to invoice offset." Compare the draft sentence against the source: the source gives only the achievement rate and the base rebate tier, and does not state the returned-goods deduction result. If the answer marks "already calculated" as pending verification and lists the missing returned-goods deduction record, the condition is met; if it directly confirms completion, the condition is not met.
- Input two follow-up turns under the same chatId. The first turn confirms the achievement rate of 92%; the second asks "then which rebate tier applies." Compare whether the second answer carries forward the first turn's interpretation. If the answer cites 92% and maps it to the base rebate tier, the condition is met; if it re-asks for the achievement rate or gives another tier, the condition is not met.
- Input a follow-up that contains no returned-goods information, and observe whether the answer writes in a returned-goods amount. If the answer clearly states that the returned-goods deduction record is missing and the final rebate amount cannot be determined, the condition is met; if it gives a specific returned-goods amount or a final amount, the condition is not met.