What the documents look like
Settlement and foreign-exchange collection work checks that funds and the underlying trade background are genuine. It does not review a full set of letter-of-credit documents. The working materials are collection receipts, trade contracts, invoices, and foreign-exchange business registration records. Entity names, amounts, currencies, and dates are spread across these materials rather than concentrated in one file. Materials update per transaction or per month, and any statement about which contract version applies should be checked against the material records for that transaction; when no version basis is supplied, the gap remains open.
Citation tracing requires that a comparison conclusion can be pointed back to specific source text. Every item the answering stage cites can be checked line by line against the original. The search node returns citations as an array, and the array length can be zero. An empty array does not mean the workflow has already errored out and stopped.
Constraints at the citation tracing step
Once retrieval output is handed to the answering stage, the answer and the cited source text must be checked item by item. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together; the model identifier must be checked first and non-matching bases excluded. The minimum relevance value is a filter threshold and does not perform permission or entity isolation. Hybrid retrieval fuses multiple result paths through RRF, and RRF is a fusion process rather than a switch that can be toggled on. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled through its settings. Optional paths shown in a flow diagram do not mean every request executes all of them.
Hypothetical drill sample: Source text one reads, "Party A Ocean Trading Co., Ltd. and Party B Hengchang Manufacturing Co., Ltd. signed goods export contract number HC-2024-017 in March 2024, amount USD 120,000, settlement method telegraphic transfer." Source text two reads, "The bank collection receipt states the payer is Hengchang Manufacturing Co., Ltd., the payee is Ocean Trading Co., Ltd., amount USD 120,000, date April 10, 2024." The draft sentence to be judged reads, "This collection has been matched to contract HC-2024-017 and settlement conversion is complete." The comparison scope is limited to these two source texts, and the completion status of settlement conversion is not stated in either source text.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select one contract base and one collection receipt base | First check that the two bases share the same vector model identifier; exclude any base that does not match |
| Search mode | Hybrid retrieval | Contract numbers are exact queries; compare semantic retrieval results after RRF fusion |
| Citation limit | Hypothetical starting point of 3000 tokens for this drill | Compare against the token volume of the contract plus receipt source texts; trim when exceeded |
| Minimum relevance | Hypothetical starting point of 0.3 for this drill | Compare against the 0-1 rerank score; candidates below the value are filtered |
| Result reranking | Enabled | Compare rerank scores against vector scores and observe ordering changes |
| Citation content output | Accept the array output | Use a paper input of [] to check whether the downstream path still runs |
Three common errors
- Treating minimum relevance as a permission filter. Input is a retrieval request against the collection receipt base; the comparison object is the minimum relevance threshold against unrelated business records; the observable result is that the threshold does not exclude unauthorized records. The check is to confirm whether the retrieval scope has already been limited to authorized knowledge bases.
- Forcing association of knowledge bases with different vector models. Input is two knowledge bases with different vector model identifiers; the comparison object is the two model identifiers; the observable result is that the identifiers do not match. The check is to verify identifiers and exclude non-matching bases, without running cross-model retrieval to compare missed recalls.
- Treating the draft sentence as a verified conclusion. Input is the draft sentence above and the two source texts; the comparison object is the settlement completion status in the draft sentence against the fields stated in the source texts; the observable result is that the source texts do not state settlement completion. The check is to record the evidence gap and mark it as pending verification.
How to verify the setup
- Input the two source texts and check the contract entity, amount, and date, plus the contract association field in the receipt. The receipt does not state a contract number, so the available materials are insufficient to confirm a match to HC-2024-017, and the association basis remains pending verification. The settlement completion portion is likewise unsupported.
- Input a paper citation array of []. Compare whether the downstream answering stage still runs. The condition is met when it runs without error. Do not claim that a fictional rule retrieval will necessarily return [].
- Input the vector model identifiers of the contract base and the collection receipt base. Compare whether the two are consistent. Only when they are consistent can they be associated together; when inconsistent, exclude them and record the actual observation, without inferring missed recalls or redundant results.