What the documents look like
Letter-of-credit document examination serves the bank's payment conditions, and the review focus is consistency across documents rather than tax matching. In store service and internal training, multi-turn prompting has to work with the issuing bank's advice, the commercial invoice, the bill of lading, the insurance policy, the draft, and the bank's document presentation records. These materials update per transaction. Within one transaction, successive turns may move through clause checking, document comparison, and discussion of discrepancies.
The material scope decides what a prompt can cite. Letter-of-credit clauses, invoice amounts, bill-of-lading dates, insurance coverage, and draft tenor are fields on different documents. A cross-document comparison needs the same transaction, the same period, and the same field. Fields the materials do not state remain unknown, and the prompt does not fill in business facts.
Constraints at the multi-turn prompts step
Multi-turn dialogue depends on history records. When a session continues, the system takes the session from the database and uses the last item in the messages array as the user question; when chatId is not passed, context is built entirely from the messages provided. In store training, the trainer needs to check the context actually carried in and distinguish the input difference between a new session and a continued session.
In a knowledge-base continuous-dialogue setting, question optimization handles reference resolution and question expansion. Whether it applies to short questions needs actual checking, and it cannot be assumed to apply to all of them. For model and length, the model's supported context and the reply limit setting need checking, with long documents and follow-up questions tested in combination. Reply format follows the downstream handoff; when json_schema is selected, the corresponding format requirement is filled in and the returned structure verified.
Retrieval results do not carry permission or entity isolation. A similarity score is a ranking reference and cannot be treated as a business-category or authorization filter. Knowledge bases require the same vector model before they can be associated together; bases with different vector models are checked by identifier and the non-matching ones excluded. For empty-result handling, a paper input of [] can check the downstream design.
Configuration and implementation checks
The settings below concern a hypothetical drill sample and list only officially described settings or implementation checks relevant to this feature and primary task.
Hypothetical drill sample: a document examiner at an exporter submits one round of checking to an internal training assistant. The supplied source text for checking is: "The issuing bank's advice states the letter-of-credit amount is USD 80,000, shipment is no later than 31 March 2025, and the bill of lading must show freight prepaid." "The commercial invoice draft amount is USD 80,000, dated 20 March 2025." "The bill-of-lading draft shows an on-board date of 2 April 2025 and freight collect." The draft sentence to be judged is: "The bill-of-lading date on this transaction is consistent with the shipment period, and the freight term is consistent with the letter-of-credit requirement." This request is pending checking and has not been authorized for external submission.
| Configuration item | Hypothetical starting point for this run | Checking method |
|---|---|---|
| AI model | Select a model that supports long context | View the maximum context and tool-calling capability shown in the dialog |
| Memory turns | Starting point set to 4 turns | This value is only a starting point for this trial, used to compare whether a follow-up cites earlier text; compare with a smaller value to observe whether earlier text is lost |
| Reply limit | Starting point set to accommodate a document summary | Compare whether the reply is truncated; compare with a longer setting for completeness |
| Temperature | Starting point set low | Rule-checking scenarios need stable output; compare with a higher value for answer variation |
| Reply format | Keep default | When a downstream process needs a fixed structure, switch to json_schema and fill in the format requirement |
| Multimodal recognition | Decided by model capability | Image, audio, and video types follow what the dialog shows; ordinary documents still need parsing into text |
Three common errors
- Treating a similarity score as an authorization or category filter. A returned score reflects ranking only and cannot exclude unauthorized records or unrelated transactions. To check, enter the same query, compare the returned list against the target retrieval scope, and observe whether records outside the scope appear.
- Treating a continued session as a new session. When chatId is passed, the system takes the database session and uses only the last message as the question, ignoring the other messages. To check, send the same question once with chatId and once without, and compare the context actually carried in.
- Treating values collected by a form as business-verified. inputForm only collects input; business verification, approval, database writes, and external synchronization need downstream implementation. To check, submit form values and compare whether a later node performs an independent verification action; if none appears, record it as pending verification.
How to verify the setup
- For the draft sentence in the hypothetical sample, input the source text and the draft sentence, then compare whether the draft sentence agrees with the bill-of-lading draft's on-board date of 2 April 2025 and freight collect. If the output confirms "consistent," that conflicts with the source text; if the output states the date exceeds the shipment period and the freight term is inconsistent, the checking condition is met.
- For the continued session, input the follow-up "Is this bill-of-lading date within the shipment period" with chatId, and compare whether the context actually carried in is the most recent turns of that session. If the output cites the shipment period from this sample, the condition is met; if the output asks for the materials again, record it as context not carried in.
- For empty-result handling, use the paper input [] to simulate a retrieval returning nothing, and compare whether the downstream node handles it as no material. If the downstream directly fabricates clause content, it does not meet the condition; if the downstream indicates the material is missing, it meets the design expectation. All of the above are paper checks and were not actually run.