What the documents look like
Refund-only dispute handling differs from ordinary return-and-refund work because there is frequently no returned parcel entering the warehouse. The evidence centre of gravity sits with dispatch and signed receipt, together with the value of the goods. For store service and internal training, the knowledge base has to carry four document types: refund-only applications, dispatch proof, signed-receipt proof, and appeal materials. After-sales supervisors and appeal specialists retrieve these documents during daily work. The brief describes the update rhythm as per order and daily. Where that rhythm does not affect this retrieval step, it can be left aside; nothing here should be read as implying batch synchronisation, full synchronisation, real-time behaviour, or transaction-triggered mechanisms.
These four document types feed three time-consuming stages: deciding whether a return is required, collecting logistics evidence, and submitting the platform appeal. Retrieval quality therefore shapes how quickly staff can assemble a defensible position. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. The output of the node is an array of quoted content whose length may be zero, which means the output path still runs even when nothing is found. Retrieval is a recall behaviour and does not perform authorisation checks; semantic recall and relevance scores do not isolate permissions or entities.
Constraints at the retrieval step
Several documented conditions shape what retrieval can and cannot settle. Association of knowledge bases requires matching vector model identifiers; where identifiers differ, the non-matching base is excluded and the observation is recorded, and no cross-model retrieval is run to compare missed recalls. The reference cap is expressed in tokens, not in a top-k item count, because chunk lengths vary widely across mixed knowledge bases. The minimum relevance value is a 0-1 figure that applies only under semantic retrieval or when result reranking is in use; it is a filtering threshold, not a final ordering rule. Question optimisation performs coreference resolution and question expansion before retrieval, and it adds one model call, so it typically lengthens the overall run. Rerank requires a model and must be enabled through settings; it produces a 0-1 score used for a second ordering pass. Hybrid retrieval combines vector and full-text results through RRF, and RRF is a fusion process rather than a switch. Image retrieval requires an image input and matching model capability; the optional paths shown in the overall flow diagram are not executed on every request.
A relevance score is not a category, entity, or permission filter. A similarity setting does not exclude unrelated businesses or unauthorised records. Retrieval does not grant authorisation, and forms collect values rather than perform business verification, approval, database writes, or external synchronisation.
Hypothetical paper exercise sample. The source material text reads: "The store after-sales supervisor received a refund-only application for Order A on 10 March 2026. The stated reason is damaged goods. Dispatch proof shows the goods left the warehouse on 8 March. The carrier trajectory shows signed receipt on 9 March." The draft sentence to be assessed reads: "This order has completed return into the warehouse, so the refund-only request can be approved." This sample is used to check whether retrieved content supports the draft sentence. No test, deployment, approval, or refund is claimed to have occurred.
Configuration and implementation checks
The table below ties each row to the hypothetical sample and to documented settings or implementation checks. Numeric starting values are hypothetical starting points for this exercise, not official defaults.
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Refund-only dispute document base | Confirm vector model identifiers match before associating; exclude bases with different identifiers |
| Search mode | Hybrid retrieval | Compare semantic, full-text, and hybrid retrieval against the test question; no single mode is presumed suitable for all material |
| Reference cap | Hypothetical starting point of 3000 tokens | Test with material of differing lengths to see whether content stays complete, then adjust the cap |
| Minimum relevance | Hypothetical starting point of 0.4 | Applies only under semantic retrieval or with result reranking; label relevant and irrelevant samples before adjusting the threshold |
| Question optimisation | Enabled | Used for coreference resolution and question expansion; compare the original and optimised question to check whether the subject was altered |
| Result reranking | ReRank enabled | Reranking requires a model and must be enabled through settings; it performs a second ordering of candidate text |
Three common errors
- Given the input "Order A refund-only application", a retrieval result containing dispatch proof and signed-receipt proof is treated as evidence of return into the warehouse. Compare the draft sentence "has completed return into the warehouse" against the source text, which records only warehouse dispatch and signed receipt. The source text does not state return into the warehouse, so the draft sentence is unsupported.
- Given an input of two knowledge bases with different vector model identifiers, both are associated together and retrieval is run. Compare the vector model identifiers of the bases. Where identifiers differ, exclude the non-matching base and record the observation; do not run cross-model retrieval.
- Given an image query, image search is assumed available without confirming whether the vector model supports image input or whether an image vector index was generated at ingestion. Compare model capability and ingestion index records. Where either is missing, record the evidence gap.
How to verify the setup
- Input the hypothetical source material text and the draft sentence. Compare the draft sentence "has completed return into the warehouse" against the dispatch and signed-receipt fields in the source text. Where the source text does not state return into the warehouse, the retrieval result does not support the draft sentence. The condition is satisfied when the draft sentence is marked unsupported.
- Input two knowledge bases with different vector model identifiers. Compare whether the identifiers match. Where identifiers differ, exclude the non-matching base and record the actual observation. The condition is satisfied when no cross-model retrieval is run.
- Input the empty-result paper input
[]. Check whether the downstream design accepts an empty array. The condition is satisfied when the output path runs and downstream handling treats the reference as empty. This does not assert that a rule-based retrieval always returns[].
Across all three checks, the same subject, the same period, and the same field must be compared. Only explicitly mutually exclusive values in the same field constitute a contradiction. Where a field is not stated, or where only an earlier related stage is evidenced, the result is unsupported or pending verification, not negated or conflicting. Pending review does not establish a lack of operating authority. Signature, agreement reached, and contract effectiveness are three distinct facts. A quotation validity period does not state a supply period. No authorisation or effectiveness clause should be invented to obtain a definite answer; evidence gaps remain open.