What the documents look like
Refund-only disputes frequently proceed without any returned goods reaching a warehouse, so the evidence weight sits on dispatch, delivery and item value rather than on inbound inspection. Operating reports and exception reviews must align four document families order by order: refund-only applications, dispatch proof, delivery confirmation, and appeal materials. These come from the platform after-sales backend, carrier tracking, customer service chat, and warehouse outbound records, and the source brief describes updates at per-order and daily frequency.
For citation tracing, this shape matters in three ways. First, a single order can carry several documents that disagree in status, so a citation must name which document and which field it draws from. Second, delivery confirmation and the buyer's claim can coexist without either resolving the other, which means a draft sentence that merges them overstates the record. Third, appeal materials are prepared by staff, so a citation to an appeal draft is a citation to an internal working document, not to platform findings.
Constraints at the citation tracing step
The knowledge base search node outputs a citation array, and the documented length may be zero. An empty array does not mean the workflow has raised an error and stopped; the documented behavior is that the output path still completes. The citation ceiling is expressed in tokens rather than a top-k item count, because chunk lengths vary widely across mixed knowledge bases. The minimum relevance value is a 0-1 threshold that takes effect only under semantic search or when result re-ranking is used, and it filters low-relevance results; it does not act as a permission filter or an entity isolation control. Linked knowledge bases must share the same vector model, and identifiers should be checked first so that non-matching libraries are excluded rather than jointly linked.
A relevance score is not a category, entity or permission filter, and retrieval is not authorization. A search returning a passage about an order does not establish that the operator is authorized to act on that order. Likewise, forms collect values; business verification, approval, database writes and external synchronization each need their own downstream implementation. No scheduled trigger or hidden control should be assumed.
Hypothetical paper exercise sample, marked as such and not drawn from any real record. Source sentences for comparison: "Refund-only application number T-0715; the buyer claims the goods were not received; application date 15 July 2026." "Outbound records show dispatch on 12 July 2026; carrier tracking shows delivery signed for on 14 July 2026." "In the customer service chat log, the buyer states that the signer was a gatehouse attendant." Draft sentence to be assessed: "Delivery verification for this order is complete and the platform appeal may be submitted." This draft places the tracking record alongside the buyer's claim without stating whether the gatehouse acceptance was authorized by the buyer, so the draft sentence is not supported by the supplied wording.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | What it is compared against |
|---|---|---|
| Linked knowledge bases | Select knowledge bases sharing one vector model | Check whether the selected libraries' vector model identifiers match |
| Citation ceiling | Hypothetical starting point of 3000 tokens | Compare against the actual length of a single appeal material |
| Minimum relevance | Hypothetical starting point of 0.5 | Compare against the 0-1 score produced after re-ranking |
| Search mode | Hybrid search | Compare against pure semantic and pure full-text results |
| Result re-ranking | Enabled | Compare against candidate order without re-ranking |
| Citation output handling | Review the citation array item by item | Compare against the quoted wording in the answer |
The starting values above are hypothetical for this page's exercise and are not official defaults. Each is chosen so that a comparison can be performed: the citation ceiling against real material length, the relevance value against a re-ranked score, and the search mode against alternative modes on the same question.
Distinguish four kinds of item. The search node's documented behavior includes producing a citation array whose length may be zero. Configurable fields include the linked knowledge bases, search parameters and re-ranking setting. Manual checks include verifying vector model identifiers and reading the citation array against the answer text. Business actions such as appeal submission, approval or refund execution require separate downstream implementation and are not performed by retrieval.
Three common errors
- Treating an empty citation array as a retrieval failure. Input is a paper record showing
[]as the node output. The correct comparison is whether the downstream node continues on the empty result, and the observable condition is that the downstream path completes as documented for a zero-length citation array. No claim should be made that the workflow necessarily raises an error. - Treating the minimum relevance value as a permission filter. The value filters low-relevance results only. It cannot exclude unrelated businesses or unauthorized records, and no comparison of relevance scores can establish authorization.
- Treating a delivery tracking record as equivalent to the buyer having received the goods. In the sample, the gatehouse acceptance has no stated buyer authorization, so the draft sentence lacks supporting wording. The comparison is between the draft's stated conclusion and the subjects, statuses and scope actually recorded in the three source sentences.
How to verify the setup
- Input the three source sentences and the draft sentence given above. Compare the draft against the subjects, statuses and scope recorded in the source wording. The condition is satisfied when the draft is observed to contain an authorization conclusion that the source wording does not state; in that case the draft is not supported.
- Input a paper citation array of
[]. Compare whether the downstream node continues execution. The condition is satisfied when the downstream path is observed to complete on the empty result, matching the documented statement that citation length may be zero. This check does not assert that any particular query returns an empty array. - Input two knowledge base identifiers that use different vector models. Compare whether the non-matching library is excluded. The condition is satisfied when the non-matching library is observed not to be jointly linked, matching the same-vector-model requirement. This check verifies identifiers and exclusion only; it does not run a cross-model search and does not predict missed recalls, redundant results or unstable behavior.
Each check states an input, a comparison and an observable condition. None of them has been executed, and no result is claimed. Where the source wording is silent, the finding remains unsupported or pending verification rather than being recorded as a contradiction. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination.