What the documents look like
Counter service and internal training for FX receipt, settlement and the trade directory work from four recurring materials: receipt vouchers for incoming foreign exchange, trade contracts, invoices, and FX business registration records. In the citation tracing step, each conclusion shown to a trainee or a customer needs to point back to a specific location in a specific valid version of one of these materials. FX receipt and settlement work centers on the authenticity of the foreign exchange funds and the underlying trade background, so the parties named, the amounts, the dates and the registration status on each material are checked item by item. Materials are refreshed per transaction and monthly, which means a citation excerpt prepared by an earlier shift can point to a superseded contract version or an expired registration record. Version confirmation is required before the excerpt is used in a training explanation.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. An empty array means the search path completed without matching content; it does not mean the workflow stopped with an error. The minimum relevance value is a filter threshold that applies only during semantic retrieval or result reranking, and that value does not perform permission isolation or entity filtering. The same-vector-model requirement applies to knowledge base association and does not extend to business registration websites or other external systems. Citation content still has to be passed to the answering step and checked line by line against the source text; fluent wording does not substitute for source evidence.
Hypothetical drill sample: an internal training file for one counter contains a passage to be checked, reading: "The date on this foreign exchange receipt voucher is 15 March 2024, the payer name is Company A, the payee name is Company B, and the amount is USD 12,000. The corresponding trade contract number is C-2024-0312, the contract signing date is 10 March 2024, and the contracting parties are Company A and Company B." A draft sentence is also submitted for assessment: "This receipt of foreign exchange has completed settlement against contract C-2024-0312, and the settlement date is 18 March 2024." The settlement date and the completed settlement status in this draft sentence are not recorded in the source text. They remain items to be verified and cannot be used directly in a training explanation.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One knowledge base that has finished training | Check the knowledge base identifier and the vector model identifier, and confirm they match the vector model used by this node |
| Search mode | Hybrid retrieval | Compare against the semantic retrieval and full-text retrieval descriptions, and confirm this mode uses vector recall and full-text recall together and merges them through RRF |
| Minimum relevance | Hypothetically set to 0.5 | This value is a trial starting point; compare with the disabled state and observe whether citation content decreases |
| Citation limit | Hypothetically set to 2000 tokens | This value is a trial starting point; compare with a smaller value and observe the change in citation content length |
| Result reranking | Hypothetically enabled | Confirm a rerank model is configured; compare with the disabled state and observe changes in citation ordering |
| Question optimization | Hypothetically disabled | Compare with the enabled state and observe differences in retrieval results during multi-turn conversation |
These values are trial starting points for this hypothetical exercise and are not official defaults. Actual configuration depends on knowledge base content and test results.
Three common errors
- Treating retrieval output as an authorization basis. The knowledge base search node returns a citation array, and that array only indicates that related content was retrieved. It does not indicate that the material has passed business review or been authorized. During counter handover, retrieval results cannot replace manual checking of the original material and its valid version.
- Treating minimum relevance as a permission filter. Minimum relevance is a threshold that filters out low-relevance search results; it does not perform the function of excluding unrelated business or unauthorized records. After the value is set, retrieval can still return content from unrelated business, and the parties and permissions still require manual checking.
- Treating an empty citation array as a workflow failure. The search node output citation array may have a length of zero, and that output path still completes. When a counter training session encounters an empty array, the check covers whether the knowledge base has finished training and whether the question is too short or lacks a subject; the empty array is not by itself a system error.
How to verify the setup
- Use the source text and the draft sentence from the hypothetical drill sample. Input the information about the receipt voucher and the contract from the source text, and compare whether the two statements in the draft sentence, "the settlement date is 18 March 2024" and "has completed settlement," have corresponding records in the source text. The source text records only the receipt voucher date and the contract signing date, and does not record a settlement date or a completed settlement status. The draft sentence is therefore unsupported content and cannot serve as a confirmed conclusion in a training explanation.
- Use the handover request from the hypothetical drill sample. Assume the counter handover request reads: "Please have the next shift explain to trainees the settlement process for this receipt of foreign exchange based on the receipt voucher and contract C-2024-0312." Input this request and the source text, and compare whether the content the request asks to explain goes beyond what the source text records. The source text does not record the settlement process or the completed settlement status. The part the request asks to explain is content to be verified and requires settlement evidence or registration records before it can proceed.
- Use the empty-result scenario from the hypothetical drill sample. Assume the paper input is the citation array []. Compare whether the downstream answering step is designed to accept an empty array. If the downstream design displays the empty array as "no related content found," that behavior matches the documented statement that the search node output length may be zero. If the downstream design treats the empty array as a system failure and terminates the flow, that design does not match the documented description and requires adjustment.