What the documents look like
Ordinary platform refunds arise order by order and do not necessarily attach to a promise made in a single livestream session. The working corpus combines the after-sales back office, customer service chat, WMS receiving records, payment refund records, refund applications, return waybills, quality inspection photos, and refund approvals. These materials update per order and per day, and the same refund number carries different state fields at the application, receiving, inspection, and approval stages. Each stage field is checked separately during handover. A retrieval result shows historical text only; it does not decide who may act. The knowledge base search node can associate one or more knowledge bases that use the same vector model, and bases built on different vector models cannot be associated together. Before handover, the base identifier and the vector model identifier are checked.
Constraints at the retrieval step
A single retrieval combines semantic search, full-text search, hybrid search, question optimization, and reranking according to the input content and search parameters, then returns reference content after similarity filtering and reference-limit trimming. Reference content is output as an array whose length may be 0, and the downstream chain still runs when nothing is found. The minimum relevance value takes effect only under semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. Retrieval does not constitute authorization: a pending-review state does not mean the operator lacks permission, a quotation validity period does not state a supply deadline, and signing, reaching cooperation, and contract effectiveness are three separate facts.
Hypothetical paper exercise sample: the source text to be checked consists of three sentences. First: "Refund number R-2031, application date 4 March 2025, refund amount 268 yuan, current state pending review." Second: "Return waybill number SF-7782, receiving date 7 March 2025, WMS record state signed for." Third: "Quality inspection photo Q-556, taken 8 March 2025, note reads outer packaging damaged, main product body shows no abnormality." The draft sentence to be judged is: "This refund has completed refund approval and the refund has been paid." The sentence shares the same subject, the same refund number, and the same fields as the source text: the source approval field reads pending review, and the source does not state a refund field. The approval completion and refund completion claimed by the draft sentence are unsupported and are recorded as pending verification, not as a denial or a conflict.
Configuration and implementation checks
The table below lists documented settings or implementation check items around this sample. The values are starting points for this hypothetical exercise only and are adjusted after comparison with the annotated sample.
| Configuration item | Starting point for this hypothetical exercise and reason | Comparison and check method |
|---|---|---|
| Associated knowledge bases | Select bases holding refund applications, return waybills, inspection photos, and refund approvals | Check each base's vector model identifier, exclude those that differ, and record the actual observation |
| Search mode | Run semantic search and full-text search separately first | Compare hit differences between the two modes using number-type questions such as refund and waybill numbers |
| Reference limit | Start at a token amount covering one full refund approval text | Test with actual materials of different lengths and check whether reference content stays complete |
| Minimum relevance | Start at 0.5, as a starting point for this exercise only | Annotate relevant and irrelevant samples first, then compare the returned set after threshold changes |
| Question optimization | Enable for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether subjects such as refund numbers were altered |
| Result reranking | Enable the rerank model when candidate results are numerous | Compare ordering before and after reranking and check whether the score is used for filtering |
Three common errors
- Treating a retrieval hit as a permission basis. Input is refund number R-2031 and the draft sentence; the comparison is between the retrieved reference content and the approval field; the observed result is that the reference content contains only the pending-review state. Concluding from this that refund permission has been granted treats retrieval as authorization.
- Reading pending review as a lack of permission. Input is the first sentence of the source text; the comparison is the value of the approval field; the observed result is that the field value is pending review. Concluding from this that the subject has no operating permission confuses a review state with a permission state.
- Treating an empty reference array as a rule guarantee. Input is the paper input []; the comparison is how the downstream design receives an empty array; the observed result is that the chain runs and the reference length is 0. Claiming from this that a fictional rule always returns an empty array over-extends the documented empty-result output condition.
How to verify the setup
- For the draft sentence "This refund has completed refund approval and the refund has been paid," input the three source sentences and the draft sentence, compare the approval field and the refund field under the same refund number, and the condition is met when the approval completion and refund completion claimed by the draft sentence are both absent from the source text, with the conclusion recorded as pending verification.
- For cross-model association, input the candidate base list and their vector model identifiers, compare whether the identifiers match, and the condition is met when only matching identifiers are retained and excluded bases are recorded, without running cross-model retrieval to compare missed recalls.
- For empty-reference handling, input the paper input [] and the downstream node design, compare the reference array length with the downstream processing branch, and the condition is met when the downstream receives length 0 normally, without claiming that a fictional rule always returns [].
Missing evidence stays missing. Pending review does not demonstrate a lack of operating permission. Retrieval does not constitute authorization. Undocumented system errors and automatic business actions are not asserted.