What the documents look like
Refund-only dispute handling differs from ordinary return-and-refund work because there is usually no returned parcel entering the warehouse. The evidence therefore centres on dispatch, delivery and the value of the goods. For product and marketing copy drafting, four document types matter: the refund-only application, the dispatch proof, the delivery confirmation, and the appeal material. The supplied brief gives an update frequency of per order and daily; this page draws no batch synchronisation or real-time trigger from that.
Sources named in the brief are the platform after-sales backend, carrier tracking, customer service chat, and warehouse outbound records, with a stated confidence of broadly so. These sources mean the retrieval library holds both structured fields and free text. Whether a promotional sentence in a draft can stand depends on whether the corresponding documents record the same party, the same object and the same period.
Constraints at the retrieval step
Retrieval treats the library as the smallest search unit; a collection is only a classification layer and does not affect search results. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries with different vector models cannot be associated together, so the model identifier is checked first and non-matching libraries are excluded. Semantic recall and the minimum relevance score act on similarity only; they do not perform permission or entity isolation, and a retrieval result is not a business authorisation.
A hypothetical paper exercise follows. Source text one: a shop after-sales record states order number A-0001, product a folding storage box, buyer signed on 2 March 2025, refund-only application submitted on 5 March 2025 on the ground of outer carton damage. Source text two: carrier tracking shows the same order signed in person on 2 March, with no return record. Source text three: the warehouse outbound record states one unit of the same model was dispatched on 28 February with packaging recorded as intact. Draft sentence under assessment: this product was signed for by the buyer and no return occurred, the packaging damage is not the shop's responsibility, so no refund is granted. The sentence concerns attribution of responsibility. The materials record signing and the absence of a return; they do not record the cause of the damage or any finding of responsibility, so the sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | One after-sales copy library using the same vector model | Check each library's vector model identifier first, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Try semantic search and hybrid search separately on the sample question | Compare whether the referenced content under each mode covers the order number and the signing date |
| Reference limit | Hypothetical starting value 2000 tokens | Test with real materials of differing length for completeness, then compare with a larger value |
| Minimum relevance | Hypothetical starting value 0.4 | Label relevant and irrelevant samples first, then compare filtering against an adjacent value |
| Question optimisation | On for multi-turn follow-ups, off for a complete single sentence | Compare the original question with the optimised question and check whether the subject was altered |
| Result re-ranking | Enabled per settings when a re-rank model is available | Compare candidate ordering before and after; when not enabled, no re-rank score is introduced |
These numbers are hypothetical starting values for this exercise, not official defaults. RRF is the fusion step for hybrid retrieval results and is not a separate switch.
Three common errors
- Writing the signing record in source text two as a finding of responsibility. The input is the carrier tracking and the outbound record; the comparison object is the responsibility wording in the draft sentence; the observable result is that the materials do not record the cause of the damage, so that wording is not supported.
- Treating a retrieval hit as a basis for authorisation. The input is the referenced content returned by retrieval; the comparison object is the evidence scope available to the draft sentence; the observable result is that the referenced content contains no approval or authorisation field, so it cannot be used to conclude that a no-refund decision has been approved.
- Writing a pending-review state as an absence of permission. The input is the current state field of the refund-only application; the comparison object is the wording about operating permission in the draft sentence; the observable result is that the state field indicates a process stage only, and pending review does not show that operating permission is lacking.
How to verify the setup
- Take source text one and the draft sentence, input order number A-0001 and the signing date, and compare whether the referenced content returned by retrieval contains both that order number and the 2 March signing record. If either item is missing from the referenced content, record the gap; do not treat it as a contradiction.
- Take source text three and the packaging-state wording in the draft sentence, input the warehouse outbound record, and compare whether the packaging state field in the outbound record and the draft wording are mutually exclusive values of the same field. Record a contradiction only when the values of the same field are clearly mutually exclusive; otherwise record it as pending verification.
- Use a paper input of an empty array to check the downstream handling design, and observe whether the output chain still runs when the referenced content length is zero. This check verifies empty-result handling only; it does not claim that retrieval under any invented rule must return an empty array.
The technical facts checked for this page are that the knowledge base search node can associate one or more knowledge bases using the same vector model; that search mode should be compared across semantic, full-text and hybrid retrieval rather than assuming one mode suits all materials; that the reference limit is measured in tokens and should be tested with materials of differing length; that minimum relevance applies only in semantic search or when result re-ranking is used, with relevant and irrelevant samples labelled before adjusting the threshold; and that question optimisation performs coreference resolution and question expansion, so the original and optimised questions are compared for subject changes.