What the documents look like
Cross-border independent-store payment risk work turns on card-network disputes and receiving accounts. Order fulfillment material centered on parcel status is a different matter. In this category the material comes from Stripe or PayPal back offices, Shopify orders, carrier proof of delivery, and customer service email. Day-to-day documents include payment chargeback notices, fraudulent order lists, invoices, and logistics proof of delivery. Update frequency is daily and per order.
For the copy drafting task, retrieval and recall must answer one question: can a promotional statement in a proposed draft be found in chargeback notices, invoices, or proof of delivery with the same subject, the same period, and the same field? A retrieval hit is a candidate citation only. It does not confirm the business claim in the copy, and it does not enforce permission isolation.
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; check the model identifier first and exclude any base that does not match. Cited content is output as an array, and its length can be 0. When nothing is found, the output path still runs through.
The citation cap is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Neither setting filters by business entity or by permission. Question optimization performs coreference resolution and question expansion, and it adds one model call. Hybrid retrieval merges multiple result paths through RRF; RRF is a fusion step, not a switch. Image retrieval requires image input and matching model capability, and reranking requires a model and must be enabled in settings. These optional paths do not all execute on every request.
Hypothetical paper exercise sample: the source material reads, "This store received three chargeback notices between March 1, 2025 and March 31, 2025, covering orders A, B, and C; the carrier proof of delivery for those orders shows delivery dates from March 5 to March 20, 2025; invoices for the same period have been issued, and the amounts match the orders." The draft sentence to be judged is, "All orders from this store ship within 24 hours and provide full tracking." The review request is: determine, based on the material above, whether this draft sentence is supported. The material does not state a shipping deadline or a tracking field, so the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the payment risk base built on the same vector model | Check the vector model identifier first and exclude bases on other models; do not run cross-model retrieval comparisons |
| Search mode | Test the draft sentence with semantic retrieval first | Compare cited content for the same question against full-text retrieval and hybrid retrieval |
| Citation cap | Set to a token amount that covers one complete chargeback notice paragraph | Test with actual material of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Label related and unrelated samples first, then set a threshold | Applies only under semantic retrieval or result reranking; compare candidates before and after filtering |
| Question optimization | Enable in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when there are many candidates | Reranking produces a score from 0 to 1; filter by that score and compare against the vector score |
Three common errors
- Treating a retrieved chargeback notice as support for the draft sentence. The error is ignoring the same-subject, same-period, same-field requirement; the material states delivery and invoicing only, and does not state a shipping deadline. Check: input the hypothetical material and the draft sentence above, compare the draft sentence fields against the material fields, and record the result as unsupported when the material does not state that field.
- Treating minimum relevance as a business or permission filter. The error is assuming a higher threshold excludes unrelated businesses or unauthorized records. Check: input the same question and adjust the threshold, compare how candidate content changes, and observe that the result reflects relevance ordering only, with no entity or permission exclusion.
- Forcing knowledge bases on different vector models into one association. The error is ignoring the same-vector-model requirement. Check: input two knowledge base identifiers on different vector models, compare whether the identifiers match, exclude the base when they do not, and record the actual observation without inferring missed recall or redundant results.
How to verify the setup
- Input the hypothetical material text and the draft sentence above, compare the promotional fields of the draft sentence against the material fields, and expect an unsupported result when the material does not state a shipping deadline or tracking; do not record it as a contradiction or a denial.
- Input an explicit paper empty result of [], check how the downstream design handles empty citations, and observe whether the output path runs through; do not claim that a fabricated-rule search necessarily returns [].
- Input the same question under semantic retrieval and under hybrid retrieval, compare cited content and relevance scores, and observe how ordering changes after RRF fusion; do not guarantee a first hit or full recall.
This page addresses only cross-border independent-store payment risk, copy drafting, and knowledge base retrieval and recall. Whether a specific promotional sentence in product copy is supported by the material is the question here; this is not an enterprise identity verification tutorial.