What the documents look like
Letter-of-credit document examination and presentation serves the bank's payment conditions, and the working standard is consistency across the presented set rather than tax matching. The materials handled per transaction include the letter of credit, commercial invoice, bill of lading, insurance policy and draft. These come from the issuing bank's advice, UCP600 clauses, the freight forwarder's bill of lading and the bank's presentation records. The set is refreshed once per transaction, so a retrieval library needs traceable indexes organised by document type.
During contract and transaction material comparison, the document examiner needs to recall the original wording of credit clauses, invoice amounts, bill-of-lading dates and insurance scope from the knowledge base, then check the proposed presentation field by field. Retrieved content is candidate reference material. It does not constitute a determination that the documents are consistent, and it does not constitute bank payment authorisation.
Constraints at the retrieval step
A knowledge base search node may be associated with one or more libraries that use the same vector model. A collection is used only for classification management, while search operates on the library as the smallest search unit. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, and hybrid retrieval merges the two paths through RRF. Rerank applies a second ordering to candidate text. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The reference cap controls the amount of content per search in tokens. Question optimisation performs coreference resolution and question expansion, and it adds one model call before retrieval.
Retrieval does not perform permission or entity isolation, and a similarity score does not exclude unrelated business records. Libraries built on different vector models cannot be associated together, so the vector model identifier must be checked and non-matching libraries excluded.
Hypothetical exercise sample: the knowledge base contains a letter-of-credit source fragment reading: "Issuing bank: Bank A; applicant: Company B; beneficiary: Company C; goods: galvanised steel sheet; amount: USD 120,000; latest shipment date: 30 June 2025; presentation period: within 21 days after shipment." A draft commercial invoice sentence for proposed presentation reads: "Invoice amount USD 120,000, shipment date 28 June 2025, beneficiary Company C." The pending request is: check whether the amount and beneficiary in the draft invoice match the letter-of-credit source text.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge libraries | Select libraries containing the letter of credit, invoice, bill of lading, insurance policy and draft; confirm the vector model identifier matches | The library holding the source fragment must share a vector model with the library associated to the search node |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question; no single mode is presumed suitable for all materials | Amount and beneficiary are exact fields, so full-text and hybrid recall differences can be compared |
| Reference cap | Unit is tokens; test with materials of differing length to see whether content is complete | The sample source fragment is short, so the cap must be checked for coverage of the full clause sentence |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking; label relevant and irrelevant samples before adjusting the threshold | The amount sentence and an unrelated clause sentence in the sample serve as the labelled pair |
| Question optimisation | Used for coreference resolution and question expansion; compare whether the subject is altered between the original and optimised question | The sample request already names its subjects, so any unrelated subject introduced by optimisation can be observed |
| Result reranking | Requires a rerank model enabled in settings; the rerank score is 0-1 | With few candidates in the sample, any ordering difference with and without reranking can be recorded |
For this page the reference cap starting value is a hypothetical trial value chosen because the sample source fragment is a single short clause sentence; the comparison method is to set a value, submit the sample request, and observe whether the returned reference content still contains the amount, beneficiary and shipment date fields in full. This is a trial starting point for the exercise, not an official default.
Three common errors
- Treating a retrieval score as an authorisation decision. Suppose the draft invoice amount matches the credit amount; a retrieved hit on that clause shows only that the content was recalled. It does not show that the bank has accepted the presentation or will pay. The error is substituting a similarity score for the manual field-by-field consistency check.
- Associating libraries built on different vector models. Suppose the source fragment library uses vector model A and a separate invoice library uses vector model B; the two cannot be associated together. The error is skipping the vector model identifier check and merging the searches directly.
- Treating a pending-review status as already effective. Suppose the proposed presentation request is marked "awaiting document examiner review". That status does not mean the presentation has been completed or that the bank has accepted it. The error is reading a pending record as an executed presentation fact.
How to verify the setup
- Input the sample letter-of-credit source text and the draft invoice sentence, then compare whether the retrieved reference content contains the amount, beneficiary and shipment date fields. The condition is met when the reference content shows those fields in full and consistent with the source text. If the reference output is an empty array, record it as not recalled rather than as inconsistent.
- Input the sample request and compare the candidates returned by semantic retrieval and by full-text retrieval separately. The condition is met when both modes recall the letter-of-credit source fragment and the full-text mode shows an observable hit on the amount digits. If only semantic retrieval returns the fragment, record the mode difference without inferring that full-text retrieval necessarily fails.
- Input the sample request and compare the question text before and after question optimisation. The condition is met when the optimised question still retains the subjects "invoice amount", "beneficiary" and "letter-of-credit source text". If optimisation replaces a subject with an unrelated object, record it as a deviation introduced by question optimisation and require manual checking of the original request.
Across all three checks, the material identity is the letter of credit, commercial invoice, bill of lading, insurance policy and draft set; the object is the field-level comparison of amount, beneficiary and shipment date; the stage is pre-presentation document examination under a pending-review status; the date scope is the 30 June 2025 latest shipment date and the 21-day presentation period stated in the sample; and the authorisation boundary is that retrieval output and any pending status remain candidate information and do not authorise presentation, bank acceptance or payment.