What the documents look like
This category covers foreign-exchange receipt and settlement together with trade directory records. The verification focus is the authenticity of the FX funds and the underlying trade background. A full letter-of-credit document set is not examined here. The working materials are receipt vouchers for incoming foreign exchange, trade contracts, invoices, and foreign-exchange business registration records. These materials are updated per transaction and per month.
For contract-and-transaction material comparison, the retrieval result must support a consistency judgement over three fields of the same transaction: amount, counterparty, and period. When a material is missing or a field is not stated, the item stays recorded as pending verification. Retrieval itself does not perform permission or entity isolation, and a similarity score does not exclude records the requester is not entitled to see.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. Before any comparison, check the vector model identifier of each candidate base, exclude those that do not match, and record the observation. Do not run a cross-model retrieval to compare missed recalls.
The node outputs cited content as an array, and the length can be zero. An empty result still lets the downstream chain run. Semantic retrieval depends on vector distance. Full-text retrieval suits exact queries such as numbers, models, and proper nouns. Hybrid retrieval merges the two channels through RRF, and ReRank then performs a second ordering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The citation cap is measured in tokens. RRF is a fusion step inside hybrid retrieval, not a switch to be turned on.
Optional paths shown in the overall retrieval flow are not executed on every request. Image retrieval requires an image input and the corresponding model capability. Reranking requires the model and must be enabled by configuration. Problem optimization performs coreference resolution and question expansion before retrieval; it adds a model call and therefore increases elapsed time.
A hypothetical paper exercise sample is set out here. The source material reads: "Exporter A signed a contract with overseas buyer B on 10 March 2025 for machinery and equipment, amount USD 120,000, settled by telegraphic transfer; on 18 March 2025 the bank received USD 120,000 remitted by B, and the receipt voucher shows the payer as B." The draft sentence to be assessed reads: "This receipt has been verified as payment under the contract and settlement may be processed." Comparison requires the counterparty B, the period March 2025, and the amount USD 120,000 to agree across the contract and the voucher, and requires the registration record to state the authorization status of this transaction. The phrase "settlement may be processed" describes a business authorization action. A retrieval result and a collected form do not constitute authorization or execution. This request remains pending verification.
Configuration and implementation checks
| Configuration item | Relevant setting on this page | Check point |
|---|---|---|
| Linked knowledge bases | Select the receipt voucher, contract, and invoice bases that use the same vector model | Verify the vector model identifier first, exclude bases on a different model, and do not perform cross-model retrieval comparison |
| Search mode | Test semantic, full-text, and hybrid retrieval separately using counterparty name, amount, and contract number | Give priority to full-text or hybrid for number-type queries; do not assume one mode suits every material |
| Citation cap | Set the starting value for this paper exercise at 2000 tokens | This value is only the starting point for this exercise; compare content completeness using contract and invoice fragments of differing length |
| Minimum relevance | Set the starting value for this paper exercise at 0.4 | Label relevant and irrelevant samples before adjusting the threshold; the value takes effect only under semantic retrieval or result reranking |
| Problem optimization | Compare the original question with the optimized question | Check whether counterparty, period, and amount were altered; observe short questions and multi-turn follow-ups separately |
| Result reranking | Enable ReRank by configuration | The rerank score is 0-1 and can be used for filtering; confirm the model is available and enabled by configuration |
Three common errors
- Treating retrieval similarity as a means of permission or entity isolation. The mistaken behaviour is assuming the minimum relevance value excludes unauthorized records or unrelated businesses. The check is to enter one counterparty name, compare whether returned citations contain records of other counterparties, and observe whether score filtering affects ordering only.
- Linking knowledge bases built on different vector models. The mistaken behaviour is ignoring the requirement that linked knowledge bases use the same vector model. The check is to verify each base's vector model identifier, exclude those that do not match, and record the actual observation without asserting a runtime result.
- Treating a pending review request as an authorized execution. The mistaken behaviour is reading the draft sentence "settlement may be processed" as a completed approval. The check is to compare the draft sentence against the authorization field in the registration record; when the field is not stated, record it as pending verification and do not supply an effective clause.
How to verify the setup
- Enter the contract text and the receipt voucher text from the hypothetical sample, and compare the counterparty B, the period March 2025, and the amount USD 120,000. The comparison condition is met only when all three fields agree and the registration record states the authorization status. When the record does not state it, record the gap.
- Enter the draft sentence "This receipt has been verified as payment under the contract and settlement may be processed," and compare it against the authorization field found in retrieval. Observe whether an explicit authorization record exists. When no record exists, keep the item pending verification and do not treat it as completed.
- Enter a clearly paper-based empty result [], and check whether the downstream design handles it through the empty-citation chain. Observe whether the output length is zero. Do not claim that retrieval under an invented rule necessarily returns [].
These checks describe what to compare and what remains unknown. No test, deployment, approval, refund, or other business action is asserted here.