Brand repair services: citation tracing for after-sales questions

Review citation tracing for after-sales questions in Brand repair services. Check sources, input boundaries, error handling and repeatable test evidence.

What the documents look like

Brand after-sales repair work centres on warranty responsibility and physical repair outlets. The materials handled in this category are warranty cards, repair work orders, inspection reports and replacement approvals. The main data sources are the after-sales system, store orders, repair outlet feedback and the national three-guarantee catalogue, with records updated per order and daily. Order fulfilment and after-sales answering need each reply anchored to the wording of a specific material, because warranty scope, spare-part coordination and replacement or return approval all depend on whether a citation can be traced and reviewed. Material versions and validity ranges differ, so the same work order can show different states at different stages. A citation must retain the subject, object, state and validity range.

A knowledge base search node can associate one or more knowledge bases that share the same vector model, and it outputs a citation array whose length can be zero. When the citation array is empty, the output path still runs; this does not mean the workflow has errored or terminated. Search results still need to be passed to the answering step, where the reply is checked line by line against the cited source text. The minimum relevance value is a filter threshold that applies only in semantic search or when result reranking is used, and it does not perform permission or entity isolation. RRF is the fusion process for hybrid search results, not a switch that can be turned on. Image retrieval requires the corresponding image input and model capability, and reranking requires a model and must be enabled in settings; these paths are not steps that execute on every request. Knowledge bases with different vector models cannot be associated together; the identifier is checked and non-matching bases are excluded, without running cross-model retrieval to compare missed recalls.

Hypothetical drill sample: a brand after-sales centre receives work order W-2025-0417. The customer submits warranty card number BK-8871, purchase date 2 March 2025, and a fault description of "touch unresponsive at the lower-left corner of the screen." A store support agent drafts the sentence: "After inspection, this fault falls within the warranty scope and the screen can be replaced free of charge." The available source text is as follows: the warranty card states "twelve months whole-unit warranty; the screen is a warranted part"; the repair work order states "pending inspection, no inspection report issued yet"; the inspection report field is blank. The handover request reads: "Please arrange a free screen replacement within the warranty scope at the repair outlet." This request is in a pending-review state and has not been authorised for execution.

Constraints at the citation tracing step

The first constraint is that a retrieved citation shows the material exists; it does not show that a replacement approval has been completed. The drafted sentence asserts an inspection conclusion, while the inspection report field is blank, so that conclusion is unsupported by the source text. The second constraint is that an empty citation array is a valid output, not a failure signal. A downstream answering step may still be invoked with zero citations, and its handling of that case is a design question rather than an error. The third constraint is that relevance filtering and permission filtering are separate concerns. A citation can be relevant and still sit outside the authorised set for a given handover. The fourth constraint is that the pending-review state of the handover request is not an execution authorisation; a pending request is not an approval, and an offered option is not an executed action.

Configuration and implementation checks

Configuration itemHypothetical starting point for this drillReason and comparison method
Associated knowledge baseOne after-sales system knowledge base, vector model identifier recorded as M1The vector model identifier of each selected base is checked for consistency before joint association; bases with an identifier different from M1 are excluded
Search modeHybrid searchWork order numbers and fault descriptions appear together; full-text search handles the number while semantic search handles the natural-language description
Minimum relevanceStarting point set to 0.3This value is a hypothetical starting point for this drill, compared with 0.5 to observe changes in the number of cited items; it is not an official default
Citation limitStarting point set to 2000 tokensThis value is a hypothetical starting point for this drill, compared with 4000 tokens to observe the length of cited content
Question optimisationEnabledCompletes references in multi-turn follow-ups; compared with the disabled state to observe retrieval hits
Result rerankingEnabled per settingsEnabled when a rerank model is available; compares citation ordering before and after reranking

Three common errors

  1. Treating retrieval output as an authorisation basis. Retrieving the warranty card text shows only that the material exists; it does not show that replacement approval is complete. Check: input the drafted sentence and the warranty card text, compare whether "falls within the warranty scope" and "replacement approval passed" are the same field with the same value, and observe whether the two are merged into one fact.
  2. Treating an empty citation array as an error. The search node output length can be zero, and the output path still runs with an empty array. Check: input the paper citation array [] and compare whether the downstream answering step is still invoked, observing whether an undefined error assertion appears.
  3. Treating minimum relevance as a permission filter. This value only filters low-relevance results and does not isolate unauthorised records. Check: input one relevant but unauthorised work order citation, compare whether that item can still appear in the citation array before and after filtering, and observe whether it is incorrectly described as permission control.

How to verify the setup

  1. Citation source check. Input the drafted sentence from the hypothetical sample, "After inspection, this fault falls within the warranty scope and the screen can be replaced free of charge," compare the blank inspection report field with the phrase "after inspection," and observe whether the inspection report text supports that conclusion. When the inspection report field is blank, the conclusion is unsupported.
  2. Authorisation state check. Input the handover request "Please arrange a free screen replacement within the warranty scope at the repair outlet," compare the pending-review state of that request with an approved state, and observe whether an approval record exists. A pending-review state does not constitute execution authorisation, and the evidence gap must be recorded.
  3. Empty citation handling check. Input the paper citation array [] and compare whether the answering step is still triggered, observing whether the downstream design handles an empty array. This check does not claim that a fictional rule retrieval necessarily returns an empty array; it checks only the handling behaviour under the given input.