What the documents look like
In cross-border independent-store payment risk control, the working material centres on card-network disputes and receiving accounts rather than parcel status. The documents handled day to day include payment chargeback notices, fraudulent order lists, invoices, and carrier proof-of-delivery records. Each carries a different subject, object, and state. A chargeback notice points to the dispute stage of a specific transaction. A fraudulent order list points to a risk marking on an order. An invoice points to a settlement amount. A proof-of-delivery record points to a carrier delivery fact. When retrieval pulls these different subjects into one question, the returned references cross evidence types.
These materials update daily and per order. At the retrieval step, the system only returns references from content already stored. Whether content can be recalled depends on its ingestion and training state.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Verify the vector model identifier first, exclude those that do not match, and record what is actually observed. Do not run a cross-model retrieval to compare missed recalls.
A single retrieval combines semantic search, full-text search, hybrid search, question optimisation, and reranking according to the input and search parameters, then fuses them into reference content. Hybrid search merges multiple result paths through RRF. RRF is a fusion process, not a standalone switch. The reference cap is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Image retrieval requires a corresponding image input and model capability. Reranking requires an available model and must be enabled in settings. Optional paths do not mean every request executes all of them. Reference content is output as an array, and its length can be zero.
Hypothetical drill sample: the source text for checking consists of three sentences. "Chargeback notice DN-2201 states that transaction T-7781 is at the dispute handling stage and has not been adjudicated." "Fraudulent order list FO-0912 marks order O-3345 as pending manual review." "Invoice INV-5567 corresponds to order O-3345, and the amount matches the settlement statement." A draft sentence to be judged reads: "Order O-3345 has been determined to be fraudulent, and this chargeback has been adjudicated." The draft sentence writes a pending-review marking as a determination and a dispute stage as an adjudication. Both exceed the states recorded in the source text.
Configuration and implementation checks
| Configuration item | Relevant value or check on this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | One or more knowledge bases using the same vector model | Verify the vector model identifier first, exclude those that do not match |
| Search mode | Compare semantic, full-text, and hybrid search against the test question | Identifier-type questions such as T-7781 are better compared under full-text and hybrid |
| Reference cap | Unit is token; test completeness with materials of different lengths | The three sample sentences are short; the starting value for this hypothetical trial is 1500 tokens, chosen to cover several short materials, and compared with 800 to see whether content is truncated |
| Minimum relevance | Takes effect only in semantic search or when result reranking is used | Label relevant and irrelevant samples first, then adjust the threshold |
| Question optimisation | Used for coreference resolution and question expansion | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Requires an available rerank model and must be enabled in settings | The rerank score is 0-1 and can be used for filtering by score |
Three common errors
- Treating a retrieval hit as an authorisation conclusion. In the sample, "pending manual review" in FO-0912 is a list state. Retrieving that sentence only shows the content was recalled. It does not show the review is complete or that the order has been determined.
- Merging states from different fields into one fact. The dispute stage in DN-2201 and the risk marking in FO-0912 belong to different fields of different materials. Only explicitly mutually exclusive values in the same field form a contradiction. Combining across fields produces a conclusion that is not supported.
- Treating an optional retrieval path as executed on every request. The sample provides no image input, so the image retrieval path does not participate. When the rerank model and its enabled state are not confirmed, the rerank score cannot serve as a filtering basis.
How to verify the setup
The following checks use the hypothetical source text and draft sentence above. Each states the input, the comparison, and the observable result. All are paper checks and were not actually executed.
- Input the draft sentence and the three source sentences, and compare "has been determined to be fraudulent" with "pending manual review" in FO-0912. If the draft sentence states a stronger status than the source text records, judge it as unsupported and require the wording to match the recorded state.
- Input the question "At what stage is the chargeback for order O-3345?" and compare whether the recalled content includes the dispute handling stage wording from DN-2201. If the recalled sentence contains only the order marking and not the dispute stage, record it as an evidence gap and do not infer an adjudication outcome.
- Input an empty reference array [] to check the downstream design, and observe whether the downstream handles it through the empty-result branch. If the downstream treats the empty array as containing content, record that this branch was not designed for the documented condition that the length can be zero.