What the documents look like
In the trade and FX administration category, the working materials are per-transaction documents. An FX receipt slip records that funds arrived, a trade contract records what the parties agreed, an invoice records how goods were priced, and FX business registration materials record registration status. The category focuses on the authenticity of FX funds and the underlying trade background. It does not examine a full set of letter-of-credit documents, so retrieval targets fields such as counterparty names, amounts, currencies, contract numbers, and registration status.
Documents move forward transaction by transaction, and the same business may appear at different stages across different materials. A retrieved passage only shows that a wording exists in the knowledge base. It does not show that receipt has been credited, that settlement has been completed, or that directory status has changed.
Constraints at the vector indexes step
Knowledge bases joined for retrieval must use the same vector model. Check the actual model identifier and the input type first. Bases whose model identifiers differ are not joined. One source data block can be associated with multiple indexes, and the index field and the source text are checked separately. A hit on an index does not mean the source wording is complete.
Index content should be complete and focused, and retrieval coverage is weighed against index length for the sample at hand. Image retrieval also depends on whether an image description or image vector index was generated at ingestion. Uploading an image at query time does not by itself create a searchable condition. Semantic retrieval relies on vector distance, and full-text retrieval relies on keyword hits. Precise queries such as contract numbers fit full-text or hybrid paths better.
The minimum relevance value is a filter threshold. It takes effect only in semantic retrieval or when result reranking is used. It does not perform permission or entity isolation, and a retrieval result does not constitute business authorization. Hybrid retrieval merges multiple result paths through RRF. That merge is part of result combination and has no corresponding on/off switch.
Configuration and implementation checks
Hypothetical drill sample: a foreign trade enterprise submits an FX receipt slip stating "Payer: A Trading Co., Ltd.; Payee: B Import & Export Co., Ltd.; Amount: USD 120,000.00; Contract No.: HT-2024-031; Value date: 6 May 2024." In the same batch, a trade contract states "Seller: B Import & Export Co., Ltd.; Buyer: A Trading Co., Ltd.; Contract No.: HT-2024-031; Amount: USD 120,000.00; Signing date: 28 April 2024." The draft sentence to be judged is "A Trading Co., Ltd. has completed FX settlement for Contract HT-2024-031." That sentence concerns completion of settlement. The two materials show only receipt of funds and contract signing, and do not state a settlement outcome.
| Check item | Relevant setting or implementation check on this page | Relation to the sample |
|---|---|---|
| Vector model | Joined knowledge bases must use the same vector model; check the actual model identifier and input type first | Confirm one model identifier before ingesting the three materials |
| Source data and indexes | One source data block can be associated with multiple indexes; index and source text are checked separately | Receipt slip and contract form separate blocks, with indexes pointing to their source text |
| Index content | Build indexes from complete, focused information; weigh coverage against length for the sample | Contract number, counterparties, and amount enter the index text |
| Image index | Whether an image is retrievable depends on an image description or image vector index generated at ingestion | This sample uses text materials and does not involve image indexes |
| Retrieval mode | Number-type queries can use full-text or hybrid retrieval; semantic retrieval relies on vector distance | Contract No. HT-2024-031 fits a precise-hit path |
| Reference cap and minimum relevance | The reference cap is controlled by tokens; minimum relevance takes effect only in semantic retrieval or reranking | This hypothetical trial starting point is a reference cap of 2000 tokens, chosen because it can hold the source text of both materials, compared against returning only a single index passage |
Three common errors
- Treating a retrieval hit as confirmation of business status. In the sample, the receipt slip shows only arrival of funds, while the draft sentence says settlement is complete. The input is the retrieved receipt-slip text, the comparison object is the status field stated in the source text, and the observed result is that the source text contains no settlement-processing record. The sentence is not supported.
- Joining knowledge bases that use different vector models. The input is two knowledge bases with different model identifiers, the comparison object is the model identifier field, and the observed result is that the identifiers differ. Exclude the non-matching base and record the actual observation. Do not infer missed recall or redundancy from this.
- Treating minimum relevance as an entity filter. The input is similar text under the same contract number but another entity, the comparison object is the counterparty name field, and the observed result is that the relevance score carries no entity-ownership information. This setting cannot exclude unrelated entity records.
How to verify the setup
- Input the receipt slip and contract text above, and compare the status wording in the draft sentence "A Trading Co., Ltd. has completed FX settlement for Contract HT-2024-031" against the fields stated in the source text. The sentence is satisfied only if the source text contains a settlement-processing record with the same entity, the same contract number, and the same period. The current source text does not state this, so the judgment is not supported.
- Input a request joining two knowledge bases whose vector model identifiers differ, and compare the model identifier fields. Joint association is allowed only when the identifiers match. When they differ, record the actual observation and do not run cross-model retrieval to compare missed recall.
- Input a paper empty result
[], and compare the downstream reference-handling design to confirm that the empty-reference path can run through. This check does not assert that any invented-rule retrieval will necessarily return an empty array.