What the documents look like
After-sales repair and reship work aims to restore product use, and the main documents are repair applications, reship orders, inspection reports, and accessory outbound orders. These materials come from the after-sales system, the order system, repair sites, and the accessory warehouse WMS, with updates per order and per day. Product information and selection questions need to retrieve usable information from these materials and answer selection questions about repair and reship. Retrieval results affect three time-consuming steps: deciding between repair and reship, coordinating accessories, and tracking customer confirmation. Document quality directly determines vector retrieval precision. Index content needs to be complete and focused, one source data block can be associated with multiple indexes, and index content and source text are checked separately.
Constraints at the vector indexes step
Knowledge bases associated for retrieval must use the same vector model, so the actual model and input type are checked first. Image retrieval also depends on whether image descriptions or image vectors were generated at ingestion, and the query side alone is not sufficient. Hybrid retrieval merges multiple result paths through RRF, and the rerank model performs a second ordering of candidate text. The minimum relevance value is a filter threshold and does not carry permission or entity isolation. Semantic recall and relevance do not decide whether material falls within an authorized scope. After a model or index change, the same group of questions is used to compare cited source text, and vector scores alone do not support a performance commitment.
Hypothetical paper exercise sample: a brand after-sales site submitted a repair application on March 10, 2026, with source text reading "Customer reports model A-100 rice cooker, fault is no power, inspection report number JC-20260310-07, requesting replacement power board, accessory warehouse WMS shows 3 power boards in stock." The draft sentence to be judged is "This application has obtained accessory outbound authorization and the power board can be reshipped directly." The source text contains subject, object, status, and validity scope, and does not state outbound authorization or reship approval.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Hypothetical starting point and comparison method |
|---|---|---|
| Vector model | Associated knowledge bases must use the same vector model; check model identifier and input type first | Starting point for this exercise is the same model identifier, compared item by item with the actual knowledge base model identifier |
| Index content | Index and source text are checked separately; one source data block can be associated with multiple indexes | Starting point is one complete index per block, compared with the completeness of the sample source text |
| Image index | Whether image descriptions or image vectors were generated at ingestion | Starting point is no image index generated, compared with whether an image was uploaded on the query side |
| Search mode | Hybrid retrieval merges through RRF; rerank is enabled according to settings | Starting point is semantic retrieval, compared with actual hits for number and model questions |
| Minimum relevance | Takes effect only in semantic retrieval or with rerank; it is a filter threshold | Starting point is a middle threshold, compared with whether cited source text covers the sample fields |
| Citation limit | Citation content is controlled by tokens | Starting point is the tokens needed to cover the sample, compared with returned citation length |
Three common errors
- Treating a retrieval hit as authorized execution. The sample source text states only a request to replace the power board and a stock quantity, while the draft sentence says outbound authorization was obtained, which treats a retrieval result as an approval fact. The check compares whether the source text states an authorizing subject and an approval status.
- Associating knowledge bases across different vector models. The documentation states that associated knowledge bases must use the same vector model. The check verifies identifiers first and excludes non-matching ones, and does not run cross-model retrieval to compare missed recalls.
- Using minimum relevance to exclude unrelated business or unauthorized records. Relevance is a filter threshold and does not carry permission or entity isolation. The check compares whether returned source text falls within the target retrieval scope and separately checks subject and permission records.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence, and compare whether the draft sentence exceeds the authorization status stated in the source text. The source text does not state outbound authorization, so the draft sentence is judged unsupported and the evidence gap is retained.
- Input the inspection report number JC-20260310-07 from the sample, and compare hit differences between semantic retrieval and full-text retrieval. Number-type questions are better served by full-text or hybrid retrieval; observe whether cited source text contains that number.
- Input an empty citation array [], and check whether the downstream design accepts an output of length 0. The node citation output can have length 0, and the empty-result path completes; no claim is made that a fictional rule retrieval necessarily returns [].