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 documents rather than tax matching. The materials handled per transaction include the letter of credit, commercial invoice, bill of lading, insurance policy and draft, and the data updates with each transaction. For product and marketing copy drafting, retrieval draws citable wording from these materials, such as the goods description, shipment information and insurance scope. A citation that conflicts with the documents can lead to refusal, deductions or delayed collection.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The vector store uses the PostgresSQL PG Vector plugin with an HNSW index, while the MongoDB dataset.datas table stores the source data for vectors and an indexes field records the corresponding vector IDs. Retrieval recalls vectors first, then looks up the source data in MongoDB by vector ID; entries mapping to the same source data are merged and the highest vector score is kept. The output of cited content is an array whose length can be 0, and the output path still runs when nothing is retrieved.
Hypothetical paper exercise sample: the source text to be checked reads "Letter of credit number LC-HYP-001, applicant HYP Trading Company, goods description 'stainless steel cookware 500 sets', shipment date 31 March 2025, insurance policy covering all risks." The draft sentence to be assessed reads "This batch of 500 sets of stainless steel cookware has been insured against all risks by HYP Trading Company, with shipment no later than 31 March 2025." The comparison checks whether the party, quantity, risk cover and date in the draft sentence match the same fields in the source text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge base | One knowledge base using the same vector model, containing the source text above | Bases using different vector models cannot be associated together; verify the identifier first and exclude non-matching bases |
| Search mode | Test semantic retrieval and full-text retrieval separately first | Compare the two modes against the test question; do not assume one mode suits all materials |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with source materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Hypothetical starting point of 0.5 | Label relevant and irrelevant samples first, then tune the threshold; this value takes effect only in semantic retrieval or with result reranking |
| Question optimization | Off first, then compared using multi-turn follow-up samples | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Not enabled first; compared later when candidates are numerous | Reranking requires an available model and must be enabled in settings; the rerank score is 0-1 |
Three common errors
- Treating semantic recall or a relevance score as a permission or entity-isolation mechanism. Minimum relevance is a filter threshold, and retrieval does not perform authorization. Observable error: assuming that raising the relevance value excludes records belonging to unrelated companies. Check: enter the sample question above, compare whether returned citations still contain fields belonging to other parties, and observe whether the relevance setting changes the scope of parties.
- Writing RRF as a toggle switch after hybrid retrieval. RRF is the fusion process for mixed results and is existing node behavior. Observable error: inventing an "enable RRF" option in the configuration table. Check: locate the description of RRF in the documents and compare whether the configuration item list contains such a switch.
- Treating an unstated field as already negated. The source text does not state a supply period, so a draft sentence adding "supply period until 30 April 2025" is unsupported. Check: compare the draft sentence field by field against the source text and observe whether any value appears that the source text does not state.
How to verify the setup
- Enter the hypothetical sample question above and compare whether the returned citation content array contains the source text fragment. Condition met: the citation content is non-empty and contains "stainless steel cookware 500 sets" and "all risks".
- Enter a query containing only an image and compare the returned result. Condition met: if the knowledge base has no vision model configured and the vector model does not support image input, the image retrieval path produces no citation; record the actual observation without presuming a failure cause.
- Enter a question unrelated to the source materials and compare the returned citation content. Condition met: the citation content may be an empty array or may contain low-relevance fragments; the actual output after the minimum relevance setting applies is the basis, and no claim is made that an unrelated query must be empty.