What the documents look like
Outlet discount operations centre on clearing out-of-season stock. The working materials are discount approvals, slow-moving inventory lists, price tags and gross-margin worksheets. Their sources include ERP stock age, POS sales, outlet mall campaigns and historical discount records, refreshed weekly or monthly. In a cross-border, multilingual setting the same SKU batch may keep its approval form, price tag and margin sheet in different languages, so a single search has to reach the same underlying material across languages. The refresh rhythm does not change how retrieval is implemented on this page, and a retrieved item does not mean an approval has been granted.
A hypothetical paper sample for later checks: an outlet store clearance team is handing a batch of out-of-season footwear discount approvals to the next shift. The source text reads: "Approval form A-07, subject 2023 autumn/winter footwear, 320 units, status submitted pending review, submission date 11 March 2024." The draft sentence to be assessed reads: "A-07 discount approved, clearance may proceed." The draft and the source differ in the status field; the source states only pending review.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model; bases with different vector models cannot be linked together. Semantic search, full-text search and hybrid search each suit different material. Codes, model numbers and proper nouns lean toward full-text search, while natural-language questions lean toward semantic or hybrid search. The reference cap is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. Question optimisation performs coreference resolution and question expansion. Reranking requires a model and must be enabled by configuration. Image retrieval requires image input and the corresponding model capability. These paths do not run on every request. A similarity score reflects content relatedness only; it does not act as a permission or entity filter, and retrieval by itself does not complete authorisation.
Configuration and implementation checks
| Configuration item | Use on this page | How to check |
|---|---|---|
| Linked knowledge bases | Link the bases holding discount approvals, slow-moving lists, price tags and margin worksheets | Compare the vector model identifier of each base; exclude any that differ |
| Search mode | Compare semantic, full-text and hybrid search against the test question | Do not assume one mode suits all material |
| Reference cap | Measured in tokens; controls how much content each search cites | Use short and long real materials to check whether content stays complete |
| Minimum relevance | Filter threshold effective only under semantic search or result reranking | Label relevant and irrelevant samples first, then adjust the value |
| Question optimisation | Used for coreference resolution and question expansion | Compare the original and optimised question; check whether the subject was altered |
| Result reranking | Reorders candidates when a model is available and enabled | Check whether it is enabled and what score filter applies |
For the sample above, a hypothetical starting reference cap of 800 tokens is chosen because the approval form and its margin note together run longer than a single price tag line; it is compared against the actual token length of the retrieved approval text. This is a starting value for this paper exercise, not an official default.
Three common errors
- Treating a retrieval hit as a completed approval. Input: the draft sentence "A-07 discount approved". Comparison: the source status field "submitted pending review". The two values differ, so a hit on this material shows only that the content was recalled; whether the discount was approved remains unsupported.
- Linking knowledge bases with different vector models. Input: several knowledge base identifiers. Comparison: each base's vector model identifier. Where identifiers differ, the bases cannot be linked together; exclude the non-matching base and record the actual observation rather than inferring runtime behaviour after linking.
- Treating minimum relevance as a permission filter. Input: a search request that touches records outside the authorised scope. Comparison: returned content against the authorised scope. A relevance score does not separate business entities or permissions, so unauthorised content may still be recalled; the authorisation boundary must be checked separately.
How to verify the setup
- For the sample source and draft sentence, input the draft sentence and run a search. Compare whether the returned content includes the source phrase "submitted pending review". The condition is met when that phrase is returned and the draft sentence is not output as an approved fact. An empty return, or a return of only the draft sentence, is recorded as a miss.
- For multi-base linking, input each base identifier and compare the vector model identifiers. All must match before the bases can be linked together. Any mismatch means that base is excluded; do not run a cross-model search to compare missed recall.
- For empty-result handling, input a paper search result of []. Compare whether the downstream node still runs through. The condition is met when the output path carries the empty result normally. This check does not assert that any invented rule search will return [].
None of these steps has been executed; they describe expected observations and failure branches only.