What the documents look like
A brand concession inside a shopping mall accepts the mall's unified cashiering and settles on a concession-fee basis. Brand ERP, mall POS, finance systems and mall membership campaigns each keep their own records. The working documents for this category are joint-operation concession contracts, mall campaign letters, sales receipts and monthly settlement statements. Updates arrive daily and at month end. Cross-border and multilingual retrieval has to handle the same fact expressed in different languages across these materials, for example a Chinese contract clause and an English campaign letter describing the same concession-fee basis.
The retrieval object is the source text of these materials. Retrieved content is a citation candidate only. It does not constitute a reconciliation conclusion, an approval, or an authorization basis. A pending request is not an approval, and an offered option is not an execution.
Constraints at the vector indexes step
Vector model quality, data quality and retriever precision together affect recall. One source data block can map to multiple indexes. When index content is more focused, retrieval precision usually rises and retrieval coverage narrows accordingly. Image retrieval also depends on whether an image-description index or an image vector index was generated at ingestion time. Uploading an image at query time does not replace an ingestion-time index.
Knowledge bases associated for retrieval must use the same vector model. Knowledge bases using different vector models cannot be associated together. Semantic retrieval supports cross-lingual understanding, so a Chinese question can match an English knowledge point. That capability comes from model training and its precision is not stable. The minimum relevance value takes effect only under semantic retrieval or result reranking, and it is a filtering threshold. It does not perform permission or entity isolation. A relevance score is not a category, entity or permission filter. Retrieval is not authorization.
Hypothetical drill sample: the source material reads, "Party A's concession participates in the mall spring membership campaign from 1 March 2024 to 31 March 2024. During the campaign the joint-operation concession fee follows the original contract. The campaign resource slot takes effect after written confirmation by the mall operations department." The draft sentence awaiting judgment reads, "The concession's spring membership campaign has received the mall's written confirmation, and the concession fee settles at the campaign rate." The source states only that the resource slot takes effect after written confirmation. It does not state that confirmation was completed, and it does not state any campaign fee rate.
Configuration and implementation checks
| Configuration or check item | Hypothetical drill starting point | Reason and comparison method |
|---|---|---|
| Vector model identifier | Same model identifier | Before associating several knowledge bases, check the model identifier per base and exclude those that differ |
| Index content scope | Single material subject and effective period | Compare against the full material and check whether the index omits the subject or the period |
| Citation limit | Set one tokens value for this sample | This value is a drill starting point; compare it with the cited source length in the sample and adjust |
| Minimum relevance | Set one 0-1 value for this sample | This value is a drill starting point; compare it with the rerank score distribution and adjust |
| Full-text retrieval path | Check against the actual vector database type | Milvus at 2.5.16 or above uses BM25; other vector databases use MongoDB full-text search |
| Image index | Check whether an image index was generated at ingestion | Check this separately from query-time image input and confirm the index exists |
Three common errors
- Treating "has received the mall's written confirmation" in the draft sentence as a fact already stated in the source material. The input is the source material above together with the draft sentence. The comparison object is whether the source contains any statement that confirmation was completed. The observable result is that the source states only that the slot takes effect after confirmation, so the sentence is unsupported.
- Treating a semantic retrieval similarity score as a permission filtering result. The input is the same question and candidate citations. The comparison object is whether the citations fall inside the retrieval scope the requester may access. The observable result is that a high or low score does not reflect access permission, so the scope needs separate checking.
- Associating knowledge bases that use different vector models. The input is the model identifiers of two knowledge bases. The comparison object is whether the identifiers match. The observable result is that bases with different identifiers cannot be associated together, so the mismatched base is excluded and the actual observation is recorded.
How to verify the setup
- Check one: input the source material above together with the draft sentence. Compare whether the source states that written confirmation was completed and whether it states a campaign fee rate. The observable result is that neither is stated, so the draft sentence is unsupported and remains pending verification.
- Check two: input the source material above. Compare the draft sentence "the concession fee settles at the campaign rate" with the source phrase "the joint-operation concession fee follows the original contract." The observable result is that the same field carries mutually exclusive values, so the draft sentence is contradictory.
- Check three: input a clearly labelled paper-based empty result array. Compare whether the downstream citation path still runs through. The observable result is that the output path still runs through when the citation length is 0. This empty result is a paper-based setting and does not indicate that any actual retrieval returned an empty value.
This page addresses only brand concession operations, cross-border material and multilingual retrieval, and vector models and indexes. At operational handover, the authorization status of materials, actions and requests is a separate matter. Retrieval or a form does not constitute authorization or execution.