Brand concession operations: multi-turn prompts for after-sales questions

Review multi-turn prompts for after-sales questions in Brand concession operations. Check input mappings, source evidence and failure handling before use.

What the documents look like

A brand concession inside a shopping mall is settled through the mall's unified cashier, and the brand side reconciles concession fees against a joint-operation concession contract and a monthly reconciliation statement. The working materials are the joint-operation concession contract, mall activity letters, sales receipts, and monthly reconciliation statements. Data comes from the mall POS, the brand ERP, mall member campaigns, and the finance system, with daily and monthly update rhythms.

These materials set the boundary of what a multi-turn conversation may cite. The joint-operation concession contract defines the concession rate and the settlement basis. Sales receipts and monthly reconciliation statements support checks of individual and monthly amounts. Mall activity letters state participation conditions and resource arrangements. After-sales questions about refunds, exchanges, activity pricing, and concession attribution can only be answered from these materials; when the matching material is absent, the only correct outcome is to record the item as pending verification.

Constraints at the multi-turn prompts step

Multi-turn dialogue must distinguish a new session from a continued session. When chatId is empty, the platform does not use its stored context and builds context entirely from the submitted messages. When chatId is a non-empty string, the platform retrieves the session from its database and uses the last item in the messages array as the user question, ignoring the remaining messages. A follow-up therefore has to keep the same chatId, otherwise earlier turns are not carried in. The prompt should state the material scope, the answering task, and how to handle missing evidence, so the model does not write in business facts.

Hypothetical paper exercise. A brand concession manager submits an after-sales handover request with three quoted source lines. Joint-operation concession contract, clause 5: "Sales at this concession are collected by the mall's unified cashier, and the brand settles concession fees according to the agreed rate in this contract." Mall activity letter: "The activity period described in this letter runs from 1 March 2025 to 31 March 2025, and participating brands must submit sales details within 5 working days after the activity ends." Sales receipt: "12 March 2025, one item sold at a transaction price of 899 yuan, cashier party is the mall." The draft sentence to be judged is: "This sale has been settled at the activity price, and the concession fee has been confirmed." The draft treats the activity period, the transaction record, and the concession confirmation as one confirmed state. The materials support the activity period and the transaction fact. Concession confirmation lacks a monthly reconciliation statement.

Configuration and implementation checks

Configuration itemHypothetical starting point for this exerciseReason and comparison object
AI modelSelect a model that supports tool calling and multimodal capabilityCompared with a model that does not support multimodal input, to check whether image receipt reading is available
Memory turnsKeep a small number of history turns firstCompared with the documented default, to observe whether earlier turns are carried into a follow-up
TemperatureStart with a lower valueCompared with a higher value; service and rule-bound scenarios need steadier output
Response limitKeep the documented default enabled state firstCompared with a lowered value, to observe whether a long explanation is cut off early
Response formatKeep the defaultCompared with json_schema; switch only when a downstream node needs a fixed structure
Multimodal recognitionChoose image recognition according to the model's actual capabilityCompared with the disabled state, to check whether an image receipt can be read

Three common errors

  1. Treating retrieval results as an authorization basis. A knowledge-base return of a sales receipt fragment only shows the record exists. Refunds, exchanges, and concession confirmation require manual or downstream system execution. Retrieval is not authorization.
  2. Writing a pending-review state as completed. In the hypothetical exercise the draft sentence says the concession fee has been confirmed, while the materials contain only the activity letter and the receipt and lack the monthly reconciliation statement. The correct record is pending verification.
  3. Treating a similarity score as an entity or permission filter. The score reflects textual relevance and does not distinguish concession, brand, or authorization scope, so unrelated records cannot be excluded on that basis.

How to verify the setup

  1. Input the three quoted material lines and the draft sentence from the hypothetical exercise. Compare whether the draft sentence goes beyond what the materials support. If the output presents the concession confirmation as completed, the check does not pass.
  2. Input two follow-up turns under the same chatId. Compare whether the second turn references the first turn's content. If earlier context is not carried in, check whether chatId is consistent and review the memory-turns setting.
  3. Input a sales receipt image with image recognition enabled. Compare whether the output reads the image content. If the model does not support multimodal input or recognition is not enabled, record the case as uncovered and keep the gap open.