What the documents look like
After-sales repairs and reships aim to restore a product to use rather than close a case with a direct refund. The work is documented across four record types. Repair applications, reship orders, inspection reports and accessory outbound orders come from the after-sales system, the order system, repair sites and the accessory warehouse WMS. Records update per order and daily. In store service and internal training, frontline staff search these materials for the criteria that distinguish a repair from a reship, the coordination wording for accessories, and the confirmation wording used with customers.
Retrieval results support judgment. Whether a repair application is accepted, whether a reship order is executed, and whether accessories leave the warehouse are authorized actions owned by the after-sales system and the warehouse. The knowledge base search node does not carry authorization responsibility. A form collects values; business verification, approval, database writes and external synchronization each need their stated downstream implementation.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together. During review, confirm the vector model identifier first, exclude bases that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
The output of referenced content is an array, and its length can be 0. When nothing is retrieved, the output path still runs through. Semantic retrieval, full-text retrieval, hybrid retrieval, question optimization, image retrieval and reranking are optional paths inside the overall retrieval plan. Image retrieval requires image input and the corresponding model capability. Reranking requires a rerank model and must be enabled by configuration. These paths do not form a fixed sequence that every request executes.
The minimum relevance score takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation. A relevance score is not a category, entity or permission filter, and it does not exclude unrelated businesses or unauthorized records. The reference limit is measured in tokens and works together with differences in chunk length. RRF is a fusion step for combining multiple result paths; it is not a switch that is turned on.
Configuration and implementation checks
Hypothetical drill sample: a brand after-sales store receives material submitted by a repair site. The first record reads: "Repair application WX-2025-0417, product model A-100, customer reports screen flicker, receipt date 2025-04-17, current status is pending inspection." The same batch contains an accessory outbound order reading: "Outbound order CK-2025-0417, accessory code P-22, quantity 1, outbound date 2025-04-17, linked repair application WX-2025-0417." The draft sentence to be assessed reads: "This repair application has completed inspection, the accessory has left the warehouse, and a reship can be arranged directly." The retrieval scope is limited to the knowledge base holding the records above.
| Configuration item | Assumed trial starting point | Value to compare against |
|---|---|---|
| Associated knowledge bases | Associate only the base holding the repair application and outbound order records above | Check whether the vector model identifiers of each base match |
| Search mode | Run semantic retrieval first, then full-text retrieval, then hybrid retrieval | Differences in referenced content for the same test question across the three modes |
| Reference limit | Assumed starting point 800 tokens, because the records carry codes and status fields | Compare against the actual record length and observe whether content is truncated |
| Minimum relevance score | Assumed starting point 0.4, to retain status-type fragments | Compare against scores of samples labelled relevant and irrelevant |
| Question optimization | Off first, then on | Whether the subject of the original question and the optimized question stay consistent |
| Result reranking | Not enabled first, then enabled | Ranking differences between rerank scores and vector scores |
The assumed starting values above are trial points chosen for this drill. They are not official defaults. Each needs a stated reason and a comparison method, as shown.
Three common errors
- Treating retrieval or a form as authorization to execute. Input the draft sentence "a reship can be arranged directly" and compare it against the status field "pending inspection" in the repair application. The status field does not support the claim that inspection is complete. The record does not state inspection completion, so the draft sentence is unsupported.
- Associating knowledge bases that use different vector models and then inferring missed recalls or redundancy. Input two bases with different vector model identifiers, compare whether the identifiers match, and observe whether the joint association condition is met. When identifiers differ, exclude them and record the observation. Do not infer run results.
- Treating the minimum relevance score as a permission or entity filter. Input a query containing unrelated business, compare whether the returned referenced content is determined only by similarity, and observe whether the threshold changes business scope. The threshold does not perform permission filtering; authorized actions need downstream system implementation.
How to verify the setup
- Input the repair application, the linked accessory outbound order and the draft sentence. Check inspection, outbound and reship separately. "Pending inspection" and "inspection complete" are mutually exclusive values in the same stage field, so they conflict. The outbound order supports accessory P-22, quantity 1, and the stated outbound date. It cannot support inspection completion or customer reship authorization. Those two points remain to be verified.
- Input two knowledge bases with different vector model identifiers. Compare whether the identifiers match and observe whether the joint association condition is met. When they differ, exclude them and do not run a cross-model retrieval comparison.
- Input a paper-recorded empty result array []. Compare whether the downstream design handles a reference length of 0, and observe whether the output path runs through. Do not claim that a fictional-rule retrieval necessarily returns [].
These verification steps describe what to compare and what remains unknown. They do not assert that any test, deployment, approval or refund has occurred.