What the documents look like
Brand after-sales repair work is organized around warranty responsibility and physical repair sites, which separates it from platform after-sales handling that centers on refunds. The everyday document set is the warranty card, the repair work order, the inspection report and the replacement approval. These arrive from the after-sales system, store orders, repair-site feedback and the national three-guarantee catalogue. In cross-border and multilingual retrieval, one repair work order may be entered in Chinese and later retrieved through an English or local-language query, so citation tracing has to return to a specific document and a specific version.
Records update per order and per day. At the retrieval step the things under review are the selected knowledge base scope and the document version. Reference content is output as an array whose length may be zero. An empty reference array means the output path still runs; it does not mean the workflow has failed or stopped.
Constraints at the citation tracing step
Citation tracing asks the answering step to take each retrieved item in turn and compare the answer against the cited source text. Only knowledge bases that share the same vector model can be associated together; bases on different vector models cannot be associated. Semantic retrieval and the minimum relevance score are similarity operations and do not carry permission or entity isolation duties. Retrieval itself is not authorization.
Hypothetical drill sample: a brand after-sales centre receives a cross-border repair work order whose source text reads: "Work order number WX-2024-0417, product model A2, fault description charging port loose, hand-in date 17 April 2024, repair site an overseas authorized site, warranty status field blank." A separate inspection report reads: "Inspection conclusion: interface solder joint cracked, non-human damage, recommend in-warranty repair." The draft sentence to be judged is: "This work order is confirmed within warranty scope and a replacement can be arranged directly." That sentence covers a warranty scope judgement and a replacement approval. The warranty status field is blank in the source, and the inspection report gives only a repair recommendation without stating any replacement approval conclusion.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Select the warranty card base and the repair work order base on the same vector model | Only the same vector model can be associated; compared against and excluded the inspection report base on a different vector model |
| Search mode | Hybrid retrieval | Work order number and model are exact queries, fault description is a semantic query; compared against semantic retrieval alone |
| Reference limit | Trial starting point set to a token amount covering one complete work order | Compared against the average length of a single work order to avoid truncating the work order body |
| Minimum relevance | Trial starting point set to a medium value | Compared against a low value used when content is not found, to observe recall change |
| Question optimization | Enabled for multilingual follow-up scenarios | Compared against a single-turn explicit question scenario, where it can be disabled |
| Reference content output | Carries the reference array from the retrieval node | Compared item by item at the answering step; an empty array is handled as zero length |
Three common errors
- Treating a retrieval hit as a confirmed warranty scope. In the hypothetical sample the warranty status field is blank. Retrieving the work order source shows only that the document was recalled; the warranty scope judgement needs manual comparison against the inspection report and warranty card and cannot be derived from the retrieval result.
- Treating the replacement arrangement in the draft sentence as already approved. The inspection report gives only an in-warranty repair recommendation. Replacement approval belongs to a separate step, and the draft sentence goes beyond what the source states.
- Forcing knowledge bases on different vector models into a joint association. Cross-model bases cannot be associated. The vector model identifier should be checked first and non-matching bases excluded, with the actual observation recorded rather than running a cross-model retrieval to compare missed recall.
How to verify the setup
- Input the work order source and the inspection report source from the hypothetical sample, then compare the draft sentence "This work order is confirmed within warranty scope and a replacement can be arranged directly" against both sources. If the warranty confirmation and the replacement arrangement claimed by the draft sentence are both absent from the sources, the sentence is judged unsupported and must go back for the warranty status and approval conclusion to be supplied.
- Input the work order number WX-2024-0417 from the hypothetical sample and compare whether the hybrid retrieval result recalls that work order source. If the recalled content contains the work order number and fault description, the citation can be traced to that document. If it is not recalled, record a miss and check whether the data has finished training.
- Input an empty reference array and compare the design of the downstream answering step. If the answering step still runs with a zero-length reference array and returns a no-source notice, the empty-result handling matches the documented behaviour. If the answering step errors out and stops, record it as an unverified handling gap.