What the documents look like
Payment risk work for independent storefronts centres on card network disputes and receiving accounts. The working materials are chargeback notices, fraud order lists, invoices and proof of delivery. These sit in payment back offices, storefront order records, carrier delivery confirmations and customer service mail. New records appear daily and with each order. Operating reports and exception reviews need to trace a conclusion back to a specific document, so citation tracing means every quoted item maps to a checkable source document and states the effective scope of that document.
A knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs citations as an array whose length may be zero. A zero-length array means the downstream path still runs even when nothing was found. Retrieved results pass through similarity filtering and a citation token ceiling before being returned, and the minimum relevance value applies only in semantic retrieval or when result re-ranking is used. These mechanics mean citations can only present candidate material; they do not replace checking the original document.
Assumed paper exercise sample: a chargeback notice states "card network dispute number A-001, transaction date 10 March 2026, amount USD 480, reason code 10.4, status pending evidence submission." A proof of delivery states "waybill number SF-7788, delivery date 12 March 2026, signature present in the recipient signature field, delivery address matches the order." A draft sentence reads: "This order has been delivered, the chargeback reason does not hold, and the dispute can be submitted directly." The comparison goal is to check whether the waybill number in the delivery proof matches the order record, whether the delivery date falls after the transaction date, and whether the draft sentence goes beyond what the materials state.
Constraints at the citation tracing step
The search node returns an array of citations, and that array can be empty. An empty array is not a workflow failure. The minimum relevance value is a filter threshold, and it applies only under semantic retrieval or when result re-ranking is enabled. It is not a final ordering rule, and it does not act as a category, entity or permission filter. A similarity setting cannot exclude unrelated businesses or unauthorized records, and retrieval is not authorization.
Re-ranking requires a re-rank model that is present and enabled, and it produces a 0 to 1 score. Hybrid retrieval combines semantic and full-text recall through RRF, and the fusion process is not a switch that is turned on. Image retrieval requires image input and matching model capability, and the optional paths shown in the overall retrieval flow are not all executed on every request. Knowledge bases with different vector models cannot be linked together; identifiers are checked first and non-matching bases are excluded, without running a cross-model search to compare recall.
The citation token ceiling uses the documented token unit, not a guessed item count. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Submitting a chargeback defence, issuing a refund or acting on an account requires separate confirmation of authorization status; the search output does not carry that authority.
Configuration and implementation checks
| Configuration item | Assumed starting point for this exercise | Note and comparison method |
|---|---|---|
| Linked knowledge bases | One payment risk document base | Check the selected base and its document version; link bases together only after confirming the vector model identifier matches |
| Search mode | Hybrid retrieval | Compare recalled content against semantic retrieval alone, and observe whether identifier-type queries are hit |
| Minimum relevance | 0.5 | Assumed starting point; compare remaining citation count after filtering against 0.3 and 0.7 |
| Citation ceiling | 2000 tokens | Assumed starting point; compare returned content length against 1000 and 4000 |
| Result re-ranking | Enabled | Requires a re-rank model present and enabled; compare ordering before and after re-ranking |
| Citation display check | Item-by-item comparison with source text | Use the traceable sample to compare the answer against cited source text; thinking output does not replace source evidence |
Three common errors
- Treating retrieval output as an authorization conclusion. Retrieval returns candidate citations only. Chargeback defence submission, refunds and account actions require separate confirmation of authorization status. Check: input the chargeback notice and the draft sentence from the assumed sample, compare whether the draft sentence claims the dispute "can be submitted directly," and mark that wording as going beyond what the materials state if it appears.
- Treating a similarity score as an entity or permission filter. The minimum relevance value filters low-relevance results only; it does not exclude unrelated businesses or unauthorized records. Check: input a query unrelated to payment risk, observe whether returned content still includes material from other business lines, and record that the score did not perform an isolation function if such material appears.
- Treating an empty citation array as a process error. The search node output length may be zero, and an empty array is not a workflow termination. Check: use a clearly labelled paper input of an empty array against the downstream design, observe whether the downstream continues on empty citations, and record the design difference if the downstream halts.
How to verify the setup
- Input the chargeback notice and the proof of delivery from the assumed sample, compare the waybill number, delivery date and transaction date for the same subject and the same period, and observe whether the citation content maps item by item to these two materials. If a material name that was not provided appears in the citations, record it as an unidentified source.
- Input the draft sentence "This order has been delivered, the chargeback reason does not hold, and the dispute can be submitted directly," compare it against the two source texts, and observe whether the materials state that the chargeback reason does not hold or that submission can proceed directly. The materials state only the delivery fact and the pending submission status, so the conclusion in that sentence is unsupported.
- Input a query containing only the transaction date and amount, compare whether the returned citations include the reason code or status field, and observe whether missing fields are marked as unknown. If the answer fills in a field the materials do not contain, record it as beyond the citation scope.