What the documents look like
Brand after-sales repair work centers on warranty responsibility and physical service outlets, so the material base differs from platform refund handling. Four document types feed contract and transaction material comparison: warranty cards, repair work orders, inspection reports, and exchange approvals. These come from the after-sales system, store orders, service outlet feedback, and the national three-guarantees catalog. A single work order can carry the send-in date, fault description, inspection conclusion, and exchange opinion at the same time, with fields scattered across different status stages. Retrieval therefore faces multiple documents under one repair number whose authorization states do not match.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded. Search mode is compared across semantic retrieval, full-text retrieval, and hybrid retrieval for the test question, without assuming one mode suits all material. The reference cap is measured in tokens, and the minimum relevance value takes effect only under semantic retrieval or when result reranking is used. Question optimization handles coreference resolution and question expansion. Retrieval results and relevance scores do not perform permission filtering or entity isolation, and retrieval itself does not constitute authorization. An empty reference output passes through the downstream chain as an array of length zero. Image retrieval requires corresponding image input and model capability, and reranking requires a model and is enabled according to settings. The optional paths listed in the overall document flow diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: a brand after-sales center receives a paper document whose original text reads, "Send-in number A-0001, send-in date 5 March 2024, fault description screen flickering, inspection conclusion ribbon cable poor contact, exchange approval column blank, outlet signatory unsigned." A second sentence reads, "The warranty card states the whole-unit warranty period is 12 months from the purchase date, purchase date 20 November 2023." The draft sentence to be assessed is, "This order has passed exchange approval and an exchange can be arranged." This sample is for paper checking only and involves no real customer or legal identity number.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the after-sales repair base sharing the sample's vector model | Check the vector model identifier first, exclude bases on other models, and do not run cross-model retrieval |
| Search mode | Start from hybrid retrieval | The sample contains a number and a natural-language description; compare semantic, full-text, and hybrid recall |
| Reference cap | Start from a token amount that can fully hold the sample's two original sentences | Test with actual material of differing lengths to see whether content stays complete, then adjust the cap |
| Minimum relevance | Mark relevant and irrelevant sample passages first, then set the threshold | This value takes effect only under semantic retrieval or result reranking, and is not a ranking rule |
| Question optimization | Enable for follow-up questions containing coreference, disable for exact number lookups | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable according to settings when a usable rerank model is available | The rerank score is used for filtering and is merged with vector and full-text results through RRF |
Three common errors
- Treating a retrieval hit as an authorization conclusion. In the sample, the exchange approval column is blank and the signatory is unsigned; retrieving that work order cannot establish that approval is complete or that an exchange can proceed. The approval field value under the same order number should be checked, and a blank field means the conclusion is not supported.
- Associating knowledge bases that use different vector models. When the sample base and another base carry different vector model identifiers, the latter must be excluded and the observation recorded, without inventing missed recall or redundant results from the association.
- Treating a pending-review state as missing operating authority. A blank approval column only shows that the field states no conclusion; it does not establish that operating authority is absent. The evidence gap should be preserved rather than filling in authorization or effectiveness clauses.
How to verify the setup
- Input the sample's two original sentences and the draft sentence, then compare the draft's claim that exchange approval has passed against the original blank approval column and unsigned signatory. Record a contradiction only when the same order number and same field show clearly mutually exclusive values; this sample should be recorded as not supported.
- Input order number A-0001 and the fault description under semantic retrieval, full-text retrieval, and hybrid retrieval respectively, and observe the recalled content. Compare whether each route covers the sample's original text, record which route hit and which was empty, and do not guarantee a first hit or full recall.
- Input a paper empty result array and check whether the downstream chain continues with a zero-length reference. Compare how the downstream design handles empty references, without claiming that a fictional rule retrieval necessarily returns an empty array.