What the documents look like
This page covers knowledge-base retrieval and recall for product material and selection questions in the foreign-exchange receipt, settlement and trade directory category. The category differs from letter-of-credit document examination: receipt and settlement work concerns the authenticity of foreign-exchange funds and the trade background, and does not review a full set of letter-of-credit documents. The working materials are mainly receipt vouchers, trade contracts, invoices and foreign-exchange business registration records. Sources include State Administration of Foreign Exchange policy, bank online banking, enterprise financial ledgers and customs export data.
For retrieval, the practical difficulty is that fields such as counterparty name, amount, currency and registration status appear across several materials and must be matched between them. Whether those fields are complete determines whether the recalled content can support later checking. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together.
Constraints at the retrieval step
Semantic retrieval, full-text retrieval and hybrid retrieval suit different question forms. Numbers, model designations and proper nouns fit full-text or hybrid retrieval; natural-language questions fit semantic or hybrid retrieval. Hybrid retrieval merges vector and full-text results through the RRF formula, and because the merged scope is large and cannot be filtered directly by similarity, a rerank model is usually needed for a second ordering and score-based filtering. Rerank requires the model to be available and enabled according to the setting.
The minimum relevance value is a 0-1 number and takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final ordering rule, and it does not isolate entities or enforce permissions. The citation limit is measured in tokens and controls how much content a single search returns. Question optimization performs coreference removal and question expansion and adds one model call before retrieval. Image retrieval requires image input and the corresponding model capability. The optional paths listed in the documentation do not mean every request executes all of them.
Hypothetical drill sample. Source text one: "Domestic payee: Huachang Textile Co., Ltd.; receipt amount: USD 48,200.00; receipt date: 12 March 2025; transaction remark: goods payment." Source text two: "Exporter: Huachang Textile Co., Ltd.; contract no.: HC-2025-018; invoice no.: INV-0312; amount: USD 48,200.00." Draft sentence to be judged: "This USD 48,200.00 receipt of Huachang Textile Co., Ltd. corresponds to contract HC-2025-018, and the trade background has been verified." The two materials agree on counterparty, amount and period. The phrase "the trade background has been verified" is a business conclusion; the materials do not record a verification action or conclusion, so that part of the sentence is unsupported.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate bases using the same vector model; check the vector model identifier one by one | Receipt voucher and contract materials must share a model before joint association |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | The sample contains contract and invoice numbers, which suit full-text or hybrid retrieval |
| Citation limit | Unit is token; test content completeness with materials of different lengths | Hypothetical trial starting value 2000 tokens, chosen because the sample holds two material sets; compare with 1000 tokens for truncation |
| Minimum relevance | Effective only under semantic retrieval or result reranking; label relevant and irrelevant samples before adjusting | Hypothetical trial starting value 0.4, compared with 0.2 for filtering of irrelevant fragments |
| Question optimization | Coreference removal and question expansion; compare whether the subject is altered | When following up "which contract does this correspond to", check the subject remains Huachang Textile |
| Result reranking | Requires a rerank model enabled by setting; the rerank score is used for filtering | Enable when candidates are many; check whether the contract number ranks near the front |
Three common errors
- Treating minimum relevance as a permission or entity filter. Input the sample question "receipt materials of Huachang Textile", compare whether returned content contains records of other entities, and observe whether non-Huachang entries appear. A relevance threshold does not exclude unrelated entities; a separate permission implementation is required.
- Forcing joint association across vector models. Input the vector model identifiers of two knowledge bases, compare whether they match, and observe whether retrieval is usable after association. When identifiers differ, exclude that base and record the observation; do not infer missed recall or redundant results.
- Treating a business conclusion in the draft sentence as verified fact. Input the source materials and the draft sentence, compare whether "the trade background has been verified" has material support, and observe whether the materials record a verification action. When not recorded, mark it unsupported and preserve the evidence gap.
How to verify the setup
- Input the two source materials and the draft sentence from the hypothetical sample. Compare whether the subject, amount and contract number in the draft sentence match the materials, and observe whether "the trade background has been verified" has material support. Matching parts pass; the conclusion part is recorded as unsupported.
- Input the sample question and run full-text retrieval and hybrid retrieval separately. Compare whether contract number HC-2025-018 appears in the cited content, and observe whether both material sets are returned completely at a citation limit of 2000 tokens. When not returned, record it as a miss; a first-position hit is not guaranteed.
- Input the explicit paper empty result
[]. Compare whether the downstream design carries the empty-citation path, and observe whether the output path runs through. Do not claim that a fictitious rule search necessarily returns[]. Authorization status, signing and contract effectiveness are distinct facts, and retrieval results do not constitute authorization execution.