What the documents look like
Letter-of-credit examination and presentation work turns on documentary consistency. The materials that circulate are issuing-bank advices, UCP600 clauses, forwarder bills of lading, and bank presentation records, while the documents handled day to day are letters of credit, commercial invoices, bills of lading, insurance policies, and drafts. In store-service and internal-training settings, document examiners and bank international settlement staff refresh material per transaction, and retrieval targets the clause text, document fields, and presentation status. Material enters the knowledge base as a collection of files; the smallest search unit is the knowledge base, and a collection serves only to classify data and does not affect search results. Because the update frequency for this category is per transaction, the retrieval step checks cited content against the current document batch and does not assume batch synchronization of closed transactions.
A practical consequence is that a training query about a specific letter of credit must resolve against the clause text and the field values of that transaction, not against a general body of trade-finance prose. A second consequence is that the same knowledge base may hold question-answer material and document material together, so chunk lengths differ widely; this is why the citation ceiling is expressed in tokens rather than a top-k item count. A third consequence is that exact identifiers such as a letter-of-credit number behave differently from natural-language questions, which shapes the choice among semantic, full-text, and hybrid retrieval.
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, so the vector model identifier is checked first and non-matching bases are excluded; no cross-model search is executed to compare recall. The node outputs cited content as an array, and the length may be zero, which means the downstream path still runs even when nothing is retrieved.
Semantic retrieval computes distance between vectors and depends on model training quality; full-text retrieval depends on keyword hits and suits numbers, models, proper nouns, and error codes; hybrid retrieval uses both and merges them through RRF. RRF is a fusion step inside hybrid retrieval and is not presented as a standalone switch. Rerank applies a ReRank model to reorder candidates and produces a 0-1 score, and it requires the model to be available and enabled according to the configuration. The minimum relevance value is a 0-1 filter threshold that takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Question optimization performs coreference resolution and question expansion before retrieval and adds one model call, which raises overall latency.
Image retrieval is not a separate capability outside the knowledge base search chain. It requires an image input and the corresponding model capability: with a visual model, an image is converted to a description that joins text retrieval; with a vector model that accepts images, image vectors are matched directly. The overall flow diagram lists optional paths, and not every request exercises all of them.
Assumed drill sample: an internal training material for a store reads, "Letter of credit number LC-2024-0187, issuing bank advice date 12 March 2024, presentation period 21 days after the bill of lading date, current status pending review." A draft sentence under review reads, "This letter of credit has completed presentation and received payment." In checking that sentence, under the same subject, the same period, and the same field, the source text states only a pending-review status; the completed presentation and the payment described in the draft sentence are not supported.
Configuration and implementation checks
| Configuration item | Setting for this page | How to check |
|---|---|---|
| Associated knowledge bases | Select knowledge bases that use the same vector model | Compare vector model identifiers and exclude non-matching bases |
| Search mode | Compare semantic, full-text, and hybrid retrieval against test questions | Test the number and a natural-language question from the assumed sample separately |
| Citation ceiling | Unit is tokens; this assumed trial starting point is 2000 | Test content completeness with long and short material and compare against a lower ceiling |
| Minimum relevance | Takes effect only under semantic retrieval or result reranking | Label relevant and irrelevant samples first, then adjust the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a ReRank model enabled per configuration | Check model availability and enabled state |
The citation ceiling starting point of 2000 tokens is chosen for this assumed trial because the sample material mixes short field entries with longer clause text; it is compared against a lower ceiling to see whether cited content still covers the fields needed for the check. This value is a trial starting point, not an official default.
Three common errors
- Treating minimum relevance as a permission or entity-isolation mechanism. Input: a retrieval question spanning different subjects. Comparison: whether returned citations include material belonging to a non-authorized subject. Observable result: the relevance filter does not exclude unrelated business records, and the access scope must be checked separately.
- Describing RRF as an independent switch that can be turned on. Input: a hybrid retrieval configuration. Comparison: whether an RRF switch appears among the node parameters. Observable result: RRF appears as a fusion process inside hybrid retrieval and is not presented as a standalone configuration item.
- Reading a pending-review status as completed presentation. Input: the assumed sample source text and the draft sentence. Comparison: the value of the same field under the same subject and period. Observable result: the source states only pending review, and the completed presentation asserted in the draft sentence is not supported.
How to verify the setup
- Input the assumed sample source text together with the draft sentence "This letter of credit has completed presentation and received payment." Compare the presentation status field in the source text against the statement in the draft sentence. The condition is met when the draft sentence is identified as unsupported by the source text and the conclusion remains pending verification.
- Input the letter-of-credit number LC-2024-0187 from the assumed sample as the search term. Compare the cited content returned by full-text retrieval and by semantic retrieval. The condition is met when the number query hits the corresponding material and the cited content matches the source text.
- Input an empty citation array [] and check the downstream design. Compare how the downstream node handles the empty result path. The condition is met when the downstream path still runs under an empty result and does not produce fabricated cited content.
These checks describe expected observations and are not claims that a test, deployment, or approval has occurred. Where the source text does not state a system error or an automatic business action, that behavior remains unknown. Retrieval and relevance scoring do not grant authorization, and any business verification, approval, database write, or external synchronization requires its own downstream implementation.