What the documents look like
Letter-of-credit examination and presentation serves bank payment conditions, and the working standard is document consistency rather than tax matching. The materials come from issuing-bank notices, UCP600 clauses, freight forwarder bills of lading, and bank presentation records. The daily documents are the letter of credit, commercial invoice, bill of lading, insurance policy, and draft. These materials commonly carry Chinese and English clauses, reference numbers, dates, and amounts, so retrieval has to cover natural-language questions and exact-number queries in the same knowledge base. Whether a document has completed training and whether an index exists determine whether recalled content can be returned as cited content at all.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the practical step is to check the model identifier and exclude those that do not match. Semantic retrieval relies on vector distance and suits cross-language understanding, for example a Chinese question matching an English knowledge point. Full-text retrieval relies on keyword hits and suits reference numbers, model numbers, and proper nouns. Hybrid retrieval combines multiple result paths through RRF, and reranking requires a rerank model and must be enabled according to the configuration. The citation limit is measured in tokens, not in a top-k item count. The minimum relevance value applies only under semantic retrieval or when result reranking is used, and it is a filter threshold rather than a final ordering rule. Retrieval does not perform permission or entity isolation; a similarity score does not exclude unrelated businesses or unauthorized records. Empty cited content is permitted: the output is an array whose length can be zero, and that output path still completes even when nothing is found.
Hypothetical paper exercise. Material text supplied for checking: This letter of credit is subject to UCP600 and is valid until 30 September 2026; the presenting bank is the advising bank located in China. The commercial invoice must show letter-of-credit number LC-HYP-001, and the consignee on the bill of lading must be the party designated by the issuing bank. Draft sentence to be judged: Payment under this batch of documents has been confirmed by the issuing bank. The draft sentence states no date, amount, or authorizing party for that payment confirmation, so it remains unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Checking method |
|---|---|---|
| Associated knowledge bases | Two libraries using the same vector model | Compare the vector model identifiers and confirm they match |
| Search mode | Test hybrid retrieval first | Compare hit content against semantic retrieval and full-text retrieval separately |
| Citation limit | 2000 tokens | Test with letter-of-credit and bill-of-lading fragments of different lengths and check content completeness |
| Minimum relevance | 0.4 | Label relevant and irrelevant samples first, then adjust the threshold |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled according to configuration | Confirm a rerank model is available and observe ordering changes |
These values are hypothetical starting points for this exercise, chosen so that different values can be compared. They are not official defaults.
Three common errors
- Treating a retrieval hit as authorized execution. Returning cited content only shows that recall exists. It does not show that the issuing bank has paid, that documents have been presented, or that examination has passed.
- Treating minimum relevance as a permission or entity filter. The value is a filter threshold and cannot exclude unauthorized records or unrelated businesses.
- Associating knowledge bases built on different vector models. The documentation requires the same vector model; when models differ, the only step is to check the identifier and exclude the non-matching library.
How to verify the setup
- Enter the material text and the draft sentence from the hypothetical exercise, then compare whether the draft sentence contains a subject, date, and amount for the payment confirmation. The material text states none of these fields, so the draft sentence is judged unsupported.
- Enter letter-of-credit number LC-HYP-001 and observe hits under full-text retrieval and under hybrid retrieval, comparing whether the number is recalled exactly. If full-text retrieval hits and semantic retrieval does not, record that difference without inferring that recall must fail.
- Enter a question for which no answer was supplied and observe whether the cited content array can have length zero. If an empty array is returned, check whether the downstream design handles an empty result. Do not claim that a fictional rule search necessarily returns an empty array.