What the documents look like
After-sales repair and reship work aims to restore a product to use, and it does not close a case through a direct refund. The materials that matter here are repair applications, reship orders, inspection reports and accessory outbound records. They come from the after-sales system, the order system, repair locations and the accessory warehouse WMS. Updates arrive per order and daily.
Contract and transaction material comparison places the subject, object, status, period and quantity found in applications, forms and reports into one comparison frame. The retrieval step returns checkable source text. The reference output is an array and its length can be 0, so the output path still runs when nothing is found. Identifiers, models, dates and quantities are exact fields. A text fragment that reads as semantically similar does not replace field-level checking.
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 is checked first, and any knowledge base that does not match is excluded.
Semantic retrieval, full-text retrieval and hybrid retrieval are compared against the test question. No single mode is assumed to suit all material. Hybrid retrieval merges multiple result paths through RRF. Result reranking requires a rerank model and runs according to its setting.
The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold. It does not isolate permissions or separate business entities, and retrieval itself does not constitute authorization. The reference limit is measured in tokens and affects the length of content returned in one search. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval.
A hypothetical paper exercise sample follows. A brand after-sales desk receives a reship application. The source text of the reship order reads: "Reship order number BX-2026-0417, subject is the after-sales intake location, object is accessory warehouse A, requested reship quantity is 2 units, application date is 17 April 2026, current status is pending confirmation by the accessory warehouse." The source text of the accessory outbound record reads: "Outbound record number CK-2026-0418, object is accessory warehouse A, outbound quantity is 2 units, outbound date is 18 April 2026, status is outbound completed."
The draft sentence to be judged is: "The reship application has been completed for outbound by the accessory warehouse." The fields needed for the judgment are object, quantity, status and date. The outbound record shows outbound completed, while the reship order shows pending confirmation. The two status fields hold different values, so the draft sentence states pending confirmation as completed outbound and is not supported by the source text. On the authorization boundary, pending confirmation by the accessory warehouse is a pending state. It does not constitute outbound execution or completed authorization.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the after-sales material knowledge base that uses the same vector model | Check the vector model identifier first, exclude knowledge bases on different models, and do not perform cross-model association |
| Search mode | Hybrid retrieval | Compare with semantic retrieval and full-text retrieval on identifier questions and natural language questions |
| Reference limit | Hypothetical starting point of 800 tokens for this exercise | Compare with 500 and 1200 to see whether returned content is complete; this is a trial starting point, not an official default |
| Minimum relevance | Hypothetical starting point of 0.4 for this exercise | Label relevant and irrelevant samples first, then compare with 0.3 and 0.5; it takes effect only under semantic retrieval or reranking |
| Question optimization | Enabled | Compare with disabled to see whether the subject is altered after coreference resolution; suited to short questions and multi-turn follow-ups |
| Result reranking | Enabled according to its setting | Compare with disabled on candidate ordering; the rerank score range is 0-1 and a rerank model is required |
Three common errors
- Treating pending confirmation by the accessory warehouse as completed outbound. The input is the reship order status field "pending confirmation by the accessory warehouse." The comparison object is the outbound record status "outbound completed." These are different values in different fields, and writing the draft sentence as completed outbound is unsupported.
- Using the minimum relevance value to exclude unrelated businesses or unauthorized records. The input is the minimum relevance threshold. The comparison object is the reference content returned by retrieval. That threshold only filters low-relevance results. It does not carry permission isolation or entity separation, and retrieval does not constitute authorization.
- Associating knowledge bases on different vector models and then inferring missed recall. The input is the vector model identifiers of the two knowledge bases. The comparison object is whether the identifiers match. When the identifiers do not match, the association is excluded. Actual recall behavior is an unobserved item, and no run result is invented.
How to verify the setup
- Input the reship order source text from the hypothetical sample and the draft sentence "The reship application has been completed for outbound by the accessory warehouse." Compare the object, quantity, status and date fields. When the status field values differ, the draft sentence is judged unsupported, and the field difference is recorded. It is not judged as a conflict.
- Input the outbound record source text from the hypothetical sample and the same draft sentence. Compare whether the outbound object matches the application object and whether the outbound quantity matches the requested quantity. When object and quantity match and the status fields differ, the conclusion is that some fields are supported and the status remains to be verified.
- Input an explicit paper input of [] to check the downstream design. Observe whether the output path runs when the reference content array has a length of 0. Do not claim that a fabricated rule search necessarily returns []. Separately, test an identifier-type question under semantic retrieval and under full-text retrieval, and record the difference in hits. Do not guarantee a first hit or full recall.