What the documents look like
Repair and reship after-sales work aims to restore a product to use. A case closes when a repair is completed or a replacement part reaches the customer. A refund is not the default outcome. That goal changes what retrieval must return. Four material types decide whether a case goes to repair or reship, whether a part can be dispatched, and whether the customer has confirmed: repair applications, reship orders, inspection reports, and parts outbound records. In cross-border work these materials often carry fields in more than one language. Semantic retrieval supports cross-language matching, so a Chinese question can reach an English knowledge point.
The materials come from the after-sales system, the order system, repair sites, and the parts warehouse WMS. Updates occur per order and daily. Parsing and chunking must let the application, inspection conclusion, and outbound record of one repair order keep a document identity and version that can be traced back to the source text. Without that, multilingual recall can merge fragments from different stages into one supposed fact.
Constraints at the document chunking step
One set of source data can map to several vectors. A hit on any one vector recalls that data set. This gives long text a way to stay complete, and it requires that a condition, its explanation, and the content it constrains sit in the same data set. The vector score takes the highest score, and the same source data hit through several paths is merged.
Full-text retrieval suits exact queries such as numbers, models, and error codes. Semantic retrieval suits natural-language and cross-language phrasing. Hybrid retrieval merges the two paths through RRF. ReRank requires a model and must be enabled in settings. The reference limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Both are filtering and trimming conditions; neither carries permission or entity isolation. The association between a retrieval library and a vector model follows the identifier. Libraries with different vector models cannot be associated together.
Hypothetical drill sample: a brand after-sales ticket record reads in three sentences — "Ticket WX-2201, customer reports no sound in the left ear, requests repair. Inspection report JC-0917 concludes the left-ear unit is damaged and recommends replacing the part. Parts outbound record CH-0430 states the left-ear unit has been dispatched, quantity 1, status awaiting customer confirmation." The draft sentence to be judged is: "Reship for WX-2201 is complete and the customer has confirmed receipt." The check point is that the outbound record status is awaiting customer confirmation; the completion and confirmation stated in the draft are not supported by the materials.
Configuration and implementation checks
The table below stays with the sample above and lists only settings already described in the official documentation or implementation check items. The values are starting points for this hypothetical trial, used to compare against retrieval observations, and do not represent official defaults.
| Configuration or check item | Hypothetical starting point for this trial | Reason for the choice and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | The sample contains the number WX-2201 and natural-language phrasing; compare the semantic and full-text recall paths |
| Reference limit | Hypothetically set to 2000 tokens | The four material types in the sample are short; compare whether recalled content is trimmed excessively |
| Minimum relevance | Hypothetically set to 0.3 | Takes effect only in semantic retrieval or result reranking; compare candidate counts before and after filtering |
| Question optimization | Hypothetically enabled | Used in multi-turn follow-up and short-question settings; compare the retrieval terms before and after completion |
| Result reranking | Hypothetically enabled | Requires a rerank model; compare how rerank scores and vector scores affect ordering |
| Chunking check | Manual review | Check whether the inspection conclusion and outbound status sit in the same group as the ticket, keeping document identity and version |
Three common errors
- Treating "dispatched" on the parts outbound record as a completed reship. When the outbound record status is awaiting customer confirmation, dispatch is a warehouse-side action; customer confirmation and reship completion are separate facts. For the check, input the outbound record text, compare the status field with the completion stated in the draft sentence, and observe whether the status field explicitly reads confirmed.
- Treating minimum relevance as a permission filter. The value is a 0–1 filtering threshold that only affects the selection of low-relevance results. For the check, input the same question, compare the candidate sets under different thresholds, and observe whether unrelated business records can still appear in the results.
- Treating a successful upload as an updated retrieval index. After material changes, the actual training and index state must be checked. For the check, input the changed term or number, compare the retrieval results with the changed content, and observe whether the new content can be recalled.
How to verify the setup
- Using the three source sentences of the sample as input, search for "WX-2201 left-ear unit", compare whether the returned fragments contain the ticket, the inspection conclusion, and the outbound status, and observe whether the three sit in the same data set and can be traced back to the source version.
- Using the draft sentence as input, compare whether the retrieval results support "reship complete, customer confirmed", and observe whether the materials show only awaiting customer confirmation. If the result is empty, a separate explicit paper input [] can be used to check the downstream design; do not claim that a search on a fictional rule necessarily returns [].
- Using a Chinese question to search English materials, compare whether semantic recall reaches the corresponding knowledge point, and observe whether cross-language matching holds. For libraries with different vector models, check the identifier first and exclude those that do not match; do not run cross-model retrieval to compare missed recalls.
All judgments above assume the same subject, the same period, and the same field. Materials that are unstated or carry only related-stage evidence are recorded as unsupported or pending verification.