What the documents look like
Payment risk work at a cross-border standalone store revolves around card network disputes and receiving accounts. The material differs from parcel-status records used in order fulfilment. The main sources are the Stripe or PayPal back office, Shopify orders, carrier proof of delivery, and customer service email. Typical documents include payment chargeback notices, fraudulent order lists, invoices, and signed delivery records. These materials update daily and per order.
At the retrieval step, the knowledge base search node returns the chargeback notice, order list, invoice, or delivery record that matches a question, so a person can review it. A retrieval hit means the material was recalled. It does not mean a person holds an operating permission, and it does not mean a handover request has been approved. The knowledge base search node can link one or more knowledge bases that use the same vector model. Cited content is output as an array, and its length can be zero, so the output path still runs when nothing is found.
Constraints at the retrieval step
A knowledge base consists of libraries, collections, and data. The smallest search unit is the library; collections only classify and manage data. Vector storage uses a vector field in PostgresSQL with an HNSW index, while the MongoDB dataset.datas table stores the original data and an indexes field. One set of data can map to multiple vectors. When the same original data is hit through several paths, the results merge and the highest score is kept. The retrieval chain combines semantic retrieval, full-text retrieval, hybrid retrieval, question optimization, and reranking according to the input and parameters. Hybrid retrieval merges results through RRF. A rerank score falls in the range 0-1. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. The citation limit is measured in tokens. Image retrieval requires an image input and the corresponding model capability, and reranking requires a model and must be enabled by configuration. Optional paths listed in the documentation do not mean every request executes all of them.
Hypothetical paper exercise sample: a risk specialist at a standalone store submits chargeback material for review during a handover. The source text reads: "On 12 March 2025, order A-1042 received a card network chargeback notice for USD 268, with reason code goods not received. The carrier delivery record for the same order shows signed receipt by the recipient on 6 March 2025. The invoice for this order has been issued and its status is sent." The draft sentence to be judged is: "The chargeback evidence for order A-1042 is complete and can be submitted for defence directly after handover." The retrieval scope is limited to the chargeback notice, delivery record, and invoice related to order A-1042 in this standalone store's risk knowledge base. This sample is a paper exercise and does not involve a real customer or a real order.
Configuration and implementation checks
The table below lists documented settings or implementation checks relevant to this feature and this task, using the sample above. The values are hypothetical starting points for this exercise, used for comparison with observed results, and are not official defaults.
| Configuration item | Hypothetical starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link only the risk knowledge base that uses the same vector model | Knowledge bases with different vector models cannot be linked together; check the vector model identifier first, exclude non-matching ones, and record the observation |
| Search mode | Run semantic retrieval and full-text retrieval once each, then hybrid retrieval | The sample contains order number A-1042 and natural-language description; compare recall across the three modes |
| Citation limit | Set to 2000 tokens for this exercise | The chargeback notice, delivery record, and invoice together are long; compare 1000 and 3000 for whether content stays complete |
| Minimum relevance | Set to 0.3 for this exercise | Label relevant and irrelevant samples first, then compare how many irrelevant fragments remain at 0.2, 0.3, and 0.4 |
| Question optimization | Off for this exercise | The sample question has a clear subject; compare with it on for whether the subject is altered and whether time increases |
| Result reranking | Enabled for this exercise | Candidates are many and the question is complete; compare the 0-1 score and ordering before and after reranking |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the delivery record being recalled only shows the material exists. It does not show the handover request was approved, and it does not show a defence was submitted. A pending review state does not prove a lack of operating permission. Signing, reaching agreement, and contract effectiveness are three distinct facts.
- Treating optional paths as executed on every request. The sample uploads no image, so the image retrieval path does not trigger. Image description retrieval is unavailable when no vision model is configured. Paths listed in the documentation do not mean every request executes all of them.
- Treating the relevance threshold as a permission or entity filter. Minimum relevance is a filter threshold. It does not exclude unrelated businesses or unauthorized records, and retrieval does not carry authorization duties. Knowledge bases with different vector models cannot be linked together; check identifiers only, exclude non-matching ones, and record the observation.
How to verify the setup
- Enter the three source passages and the draft sentence from the sample, and run one search within the limited retrieval scope. Compare whether the recalled content includes the chargeback notice, delivery record, and invoice, and whether the draft sentence claim that "evidence is complete" is supported by the source text. The source text states only the notice, the signed receipt, and the invoice status. It does not state the other conditions required to submit a defence, so that conclusion is unsupported and the evidence gap should be recorded. Observe the failure branch: if recall returns an empty array, use an explicit paper input of [] to check the downstream design, and do not claim that a fictional rule makes retrieval always return [].
- Enter the same question and run it under semantic retrieval, full-text retrieval, and hybrid retrieval. Compare the hit result for order number A-1042 against the hit result for the natural-language description. Hybrid retrieval merges multiple result paths through RRF and is not written as an on/off switch. Observe whether any single path misses; a miss on one path does not mean the other modes necessarily fail.
- Enter the same question and compare the retrieval terms with question optimization off and on. Check whether the subject, order number, and date are altered. Question optimization performs coreference resolution and question expansion, and it adds one model call and overall time. Observe the failure branch: if the subject changes after optimization, record that result and keep the original question result, and do not conclude from it that optimization necessarily fails.
All of the above are paper review steps that were not actually executed, and no test, deployment, or approval is claimed.