What the documents look like
This page addresses source citation and traceability for product information and selection questions in the foreign-exchange settlement and trade directory category. The category concerns the authenticity of FX funds and trade background; it does not review a full set of letter-of-credit documents. The materials handled here are inward remittance receipts, trade contracts, invoices, and FX business registration records. The update rhythm is per transaction and monthly.
For citation tracing, the version of a material directly determines whether a retrieved citation is usable. A retrieved passage must be traceable to a specific material and its valid scope. A remittance receipt and a contract can share a contract number and still describe different stages: receipt of funds, delivery obligations, or completion of delivery. The citation task is to keep the cited sentence within what the source material actually states.
Constraints at the citation tracing step
The knowledge base search node outputs cited content as an array, and its length can be zero. A zero-length array means the downstream path still runs; it does not mean the workflow has errored or terminated. The node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together, so the identifiers must be checked and non-matching ones excluded before retrieval.
Minimum relevance is a 0–1 value that filters low-relevance search results. It takes effect only in semantic retrieval or when result reranking is used. It is a filter threshold, not a permission or entity-isolation control. RRF is the fusion step for hybrid retrieval results; it is not a separate switch. Thinking budget belongs to model thinking configuration and is unrelated to the citation token cap. Image retrieval requires image input and the corresponding model capability; reranking requires a model and must be enabled by configuration. Optional paths in the overall retrieval flow do not mean every request executes all of them.
A hypothetical paper exercise sample: Material text one reads, "On 10 March 2025, Company A received USD 120,000 from overseas Company B under contract number A-001; receipt number S-20250310-01." Material text two reads, "Contract number A-001 provides for delivery in two batches; the first batch delivery deadline is 30 April 2025." The draft sentence to be assessed is, "This inward remittance corresponds to full delivery of all goods under contract A-001." The assessment must use the same subject, the same period, and the same field. Material text two states only the first-batch deadline and does not state that all delivery was completed. The draft sentence is unsupported and the evidence gap must be recorded.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Only knowledge bases using the same vector model | Check identifiers first, exclude different-model ones, and do not run cross-model retrieval comparisons |
| Search mode | Hybrid retrieval | Compare against semantic retrieval and full-text retrieval on whether retrieved passages cover the sample materials |
| Minimum relevance | 0.35 (hypothetical starting point) | Compare with 0.2 and 0.5 on whether filtered passages contain the sample source text |
| Citation cap | 3000 tokens (hypothetical starting point) | Compare with 1500 and 6000 on whether the sample source text enters the citation intact |
| Result reranking | Enabled | Compare with disabled on the ranking position of the sample source text; a model must support it |
| Response format | Keep default | Compare with json_schema; change only when a fixed structure is required downstream |
Three common errors
- Treating a retrieval hit as authorization or business confirmation. Input: the sample inward remittance receipt. Compare the retrieval result with the authorization state. Observable result: the hit passage shows only that the material exists; it does not show that settlement or crediting has been approved.
- Treating minimum relevance as an entity-isolation measure. Input: the sample contract. Compare results before and after filtering. Observable result: the threshold affects relevance ordering only; it does not exclude unrelated businesses or unauthorized records.
- Treating an empty citation array as a workflow failure. Input: a paper-level empty array. Compare the downstream node design. Observable result: the path still runs, and the downstream handling of an empty result must be checked.
How to verify the setup
- Input sample material text one and the draft sentence. Compare the draft sentence "all goods have been delivered" with material text two's "first batch delivery deadline." Observable result: the source does not state full delivery, so the sentence is unsupported and the evidence gap is recorded.
- Input two knowledge base identifiers that use the same vector model and one identifier that uses a different model. Compare the association list. Observable result: the different-model identifier is excluded, no cross-model retrieval is executed, and the actual observation is recorded.
- Input a paper-level empty array []. Compare the downstream handling design. Observable result: the path runs and the downstream processes the empty result accordingly; no claim is made that a fictional-rule retrieval necessarily returns an empty array.