What the documents look like
Outlet discount operations aim to clear past-season inventory rather than maintain a full-price new-arrival image. The working documents are discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. In cross-border and multilingual retrieval these materials frequently carry price tags in several languages, cross-border channel discount terms and margin calculations in different currencies, so retrieval has to match meaning across languages.
A semantic search relies on vector distance and can match a Chinese question to an English knowledge point. A knowledge base selected for association must use the same vector model as the others, so the actual model and its input types are checked first. One original data block can be linked to several indexes, and the index content is checked against the source text separately. Image retrieval depends on whether an image description or image vector was generated at ingestion time. An RRF merge combines results from several recall paths and is one stage of the retrieval chain. The minimum relevance value applies only during semantic search or after result reranking, and it is a filter threshold. The reference cap is measured in tokens.
A hypothetical paper case, labelled as such, provides the material for this page: on 10 March 2025 a channel clearance group at an outlet store submits a discount approval form. The source text reads: "This form covers 2024 autumn and winter slow-moving styles, discount range fifty to seventy percent, applicable stores are three East China outlet stores, approval status is pending review by the channel lead." A draft sentence reads: "This discount approval has been granted and the price tag change can be executed." The source shows only a pending status, so the draft sentence treats a pending item as granted, which the source does not support.
Constraints at the vector indexes step
The knowledge base selected for association must use the same vector model. The model identifier and its input types are checked, and a different model is excluded from joint association. One original data block can carry several indexes, and each index is compared with the source text rather than assumed to cover it. Image retrieval depends on an index created at ingestion; uploading an image at query time does not create that index. RRF is a merge stage for mixed retrieval results, not a switch that is turned on. The minimum relevance value is a filter threshold that applies only in semantic search or after reranking, and it does not isolate businesses, categories or permissions. Retrieval is not authorization. The reference cap is counted in tokens, which differs from a top-k item count. A relevance score is not a category, entity or permission filter, and a similarity setting does not exclude unrelated businesses or unauthorized records. Where a field is not provided in a document, that field remains unknown.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this case | How to check |
|---|---|---|
| Vector model | Same as the associated knowledge base | Compare model identifier and input types; exclude a different model |
| Index content | Built from the discount terms source text | Compare against the source sentence by sentence and confirm coverage |
| Image index | Assumed to have an image vector generated | Check the ingestion record; query-time upload does not substitute |
| Search mode | Semantic search plus full-text search | Observe numbered terms and natural-language questions separately |
| Minimum relevance | Hypothetical trial start at 0.3 | Compare against the cited source text and adjust step by step |
| Reference cap | Set by tokens | Observe whether returned reference length meets the check need |
The value 0.3 is a hypothetical trial start for this page, chosen because it sits in the middle of the 0–1 range, and it is compared with higher and lower values by observing changes in the cited source text. It is not an official default. The image index row is an assumption for this case, not an observation. Fields not listed in the source document, such as a specific error code, remain unknown.
Three common errors
- Treating a pending review as granted. The discount approval source text states that the status is pending review by the channel lead, while the draft sentence states that approval has been granted. The two values occupy the same status field and are mutually exclusive, so the draft sentence is not supported by the source.
- Associating knowledge bases that use different vector models. When two knowledge bases carry different vector model identifiers, the identifiers do not match, so the non-matching one is excluded. No cross-model retrieval is run to compare missed recalls.
- Treating a query-time image upload as an existing image index. Whether an image can be retrieved depends on whether the corresponding index was generated at ingestion time. A query-time upload does not constitute an ingestion index, and the missing record is recorded as a gap.
How to verify the setup
- Input the discount approval source text and the draft sentence above, and compare the approval status field. The source value is pending review and the draft value is granted. Because the two values are mutually exclusive in the same field, the draft sentence is not supported. This is a paper comparison, not a completed test.
- Check the vector model identifiers of the two knowledge bases. Joint association is possible only when the identifiers match. When they differ, exclude the non-matching base and record the actual observation. Do not infer a runtime result from the exclusion.
- Check the image ingestion record for an image vector index. When the record exists, the image can take part in image vector retrieval. When it is missing, record the gap and do not claim that a query-time upload replaces it.
The empty-reference output remains an array whose length can be zero, so a downstream design can be checked against an explicit paper input of an empty array. Whether any particular query returns an empty array is not asserted here. Optional paths shown in the overall flow diagram are not executed on every request: image retrieval requires a corresponding image input and model capability, and reranking requires a model and its configured enablement. Where a document does not state a system error or an automatic business action, no such behavior is asserted.