What the documents look like
Channel receivables collection happens after credit sales, and the work centres on credit usage and cash recovery. The everyday materials are the receivables ageing table, the dunning letter, the credit agreement and the shipment freeze notice. Data comes from the finance system, bank receipts, dealer orders and sales contracts. Brand finance staff, sales back-office staff and dealer owners all take part, and errors can produce bad debt, a cut-off in channel supply and damage to sales targets.
Order fulfilment and after-sales Q&A concentrates on three question types: whether a specific payment has arrived, whether a batch of orders sits inside the credit limit, and whether shipment has been frozen after the limit was exceeded. Each judgement depends on the stated condition of the ageing table, the bank receipt, the contract and the freeze notice. Where any one of these is missing, the only correct record is pending verification.
Constraints at the multi-turn prompts step
The prompt has to state the document scope, the answering task and the handling of missing evidence. The model must not write in business facts that were not supplied. It must not infer a credit conclusion from a similar company name, and it must not treat a retrieved ageing record as confirmation that repayment arrived. For history, check how many context turns are actually carried in, and separate the input source of a new session from that of a continued session. A continued session relies on chatId to pull the session from the database, uses the last item of the messages array as the current question, and ignores the remaining messages.
Question optimisation applies to knowledge-base continuous conversation. Check whether reference resolution and question expansion act on the current follow-up, and do not assume that every short question qualifies. For model and length, check the selected model's maximum context and the reply limit setting, then test whether a long ageing table combined with follow-ups is truncated. Choose the reply format from what comes downstream. Ordinary Q&A keeps the default. Where a later flow needs to read a fixed structure, selecting json_schema requires filling in the corresponding format requirement and verifying the returned structure.
Hypothetical paper exercise: the source text reads "Credit agreement number XS-2024-017: Party A grants Party B a credit-sale limit of RMB 800,000, valid until 31 December 2024." and "Shipment freeze notice number DJ-2024-033: because Party B's ageing exceeds 90 days, shipment is suspended from 15 November 2024." The draft sentence under review is: "Party B's current credit limit is exhausted and over-limit shipment has been frozen; please remit payment immediately." That sentence merges two distinct facts, limit exhaustion and ageing overrun. The materials do not provide the used balance of the limit, and they do not state a causal link between the freeze and the payment demand.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Reason and comparison method |
|---|---|---|
| AI model | A model shown as supporting tool calls | Order fulfilment Q&A needs to query ageing and bank receipts; compare whether the model displays tool-call support |
| Memory turns | Assumed set to 6 turns | Collection follow-ups usually span several turns; compare whether earlier dialogue is still referenced and how the response changes |
| Reply limit | Assumed switched on at a medium length | Ageing explanations need a complete paragraph; compare whether replies are truncated early |
| Temperature | Assumed set low | Fulfilment and freeze judgements require stability; compare whether repeated answers to the same question stay consistent |
| Reply format | Keep the default | No downstream flow currently reads a fixed structure |
| Multimodal recognition | Enable image recognition only if the model supports it | Bank receipt screenshots may arrive as follow-up input; compare whether image content is read once enabled |
Three common errors
- Treating a retrieved ageing record as confirmation that repayment arrived. The input is a knowledge-base retrieval hit on the ageing table; the comparison object is the bank receipt text. The arrival condition is met only when the receipt states the matching amount and date. A retrieval hit is not evidence of arrival.
- Extending the suspension wording in a shipment freeze notice into termination of the contract. The input is the freeze notice text; the comparison object is the credit agreement and contract clauses. The freeze notice does not state terminating effect, so that conclusion is unsupported.
- Treating a repayment plan awaiting review as approved and in execution. The input is the submitted plan record; the comparison object is the completed approval record. A submitted status does not show approval, and the item should be recorded as pending verification.
How to verify the setup
- Take the credit agreement text and the draft sentence from the hypothetical exercise above. Submit the draft sentence and compare whether it contains the limit-exhausted wording that the materials do not state. If that wording appears, the prompt has not constrained the model from writing in business facts.
- Take the freeze notice text from the hypothetical exercise. Submit the follow-up "Does the freeze amount to termination of the contract?" and compare whether the answer cites the suspension scope stated in the notice. If the answer gives a termination conclusion directly, the missing-evidence handling is not in effect.
- Take the submitted repayment plan record from the hypothetical exercise. Submit "Has this plan been approved?" and compare whether the answer separates submitted status from approved status. If the answer confirms approval, the pending-review status has not been preserved.