What the documents look like
Payment risk control materials centre on card network disputes and receiving accounts. The daily working set includes payment chargeback notices, fraudulent order lists, invoices and proof of delivery signatures. These materials update daily and per order, and within one period a single order can carry a dispute status and a funds status at the same time. Operating reporting and exception review need the same order, the same period and the same field placed side by side for comparison. The retrieval step is responsible for pulling the matching materials out of the knowledge base.
This differs from parcel-status order fulfilment work, which tracks shipment state. Here the review question is whether a dispute stage, an invoice fact or a delivery signature actually supports a proposed statement or a handover request. Retrieval supplies candidate text; it does not decide business status.
Constraints at the retrieval step
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; check the model identifier first, exclude those that do not match, and record what was observed. Do not run a cross-model search to compare missed recalls.
Search mode is compared across semantic, full-text and hybrid retrieval for the same test question; no single mode is assumed to suit every material type. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimisation performs coreference removal and question expansion; compare the original question with the optimised question and check whether the subject was altered. Quoted content is output as an array and its length can be zero, so the output path still runs when nothing is found. Similarity and relevance scores do not carry permission or entity isolation duties.
Hypothetical paper sample. Material text one: "On 10 March 2025, order A-1001 received a card network chargeback notice; the dispute stage is pending review; the amount involved is 100 US dollars." Material text two: "On 10 March 2025, an invoice for order A-1001 was issued; the amount is 100 US dollars." Draft sentence to be judged: "The chargeback for order A-1001 has been fully handled." The subject needed for the judgement is order A-1001, the object is the chargeback notice, the period is 10 March 2025, and the status field is the dispute stage. A pending-review value does not show that operating permission is absent, and it does not amount to completed handling. The draft sentence is unsupported and should be recorded as awaiting verification.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One payment risk control knowledge base | Check the vector model identifier of each base first, exclude those that differ, then decide whether to add it |
| Search mode | Hybrid retrieval | Run the same test question under semantic, full-text and hybrid modes and compare the quoted content |
| Citation limit | Hypothetical starting point of 2000 tokens | The sample mixes notices and invoices of differing length; compare with 1000 and 3000 for content completeness |
| Minimum relevance | Hypothetical starting point of 0.3 | Label relevant and irrelevant samples first, then compare filtered results against 0.2 and 0.4 |
| Question optimisation | Enabled | Compare the original and optimised question against the disabled state, checking whether the subject was altered |
| Result reranking | Enabled according to settings | Compare candidate ordering with the disabled state, checking how the rerank score works with minimum relevance |
Three common errors
- Treating a retrieval hit as authorisation. Input: the draft sentence "The chargeback for order A-1001 has been fully handled." Retrieval returns the chargeback notice and the invoice. Compare the dispute stage field in the materials and observe that "pending review" does not match "fully handled". The draft sentence is unsupported. Retrieval does not perform permission judgement; the business handling status requires separate verification.
- Associating bases across vector models. Input: the vector model identifiers of two bases. Compare whether the identifiers are the same and observe that a differing identifier means the base is excluded. Do not run a cross-model search to compare missed recalls; record the actual observation.
- Writing the RRF fusion of hybrid retrieval as a switch. Input: the search mode setting. Compare the semantic and full-text result paths and observe the ordering of the fused quoted content. RRF is a fusion process and is not configured as a standalone switch.
How to verify the setup
- Input the material text and the draft sentence from the hypothetical sample, compare the values of the dispute stage field, and observe that the material reads "pending review" while the draft sentence reads "fully handled". The draft sentence is judged unsupported and the gap is recorded.
- Input the same test question and run it under semantic, full-text and hybrid retrieval respectively. Compare the returned quoted content arrays and observe the differences in recalled content across modes. A first-position hit or full recall is not guaranteed.
- Input an explicit manual empty result
[]to check the downstream design. Compare how the downstream handles the empty array and observe whether the output path runs. Do not claim that a fictional rule search necessarily returns[].