What the documents look like
Tmall flagship store daily operations centre on brand authorization, flagship store identity, and Tmall points compensation rules. Compared with JD POP stores, these three elements carry more weight in how store service and internal training materials are written and read. The main data sources are the Tmall merchant backend, the Tmall rule centre, ERP order tables, and customer service chat records. Documents handled daily include product publishing pages, brand authorization letters, delivery commitment records, and after-sales tickets. These materials are frequently distinguished by subject, object, status, and validity scope. Retrieval from a trained knowledge base returns cited content; it does not replace authorization judgment or business execution. A handover request, a pending-review status, and a signed document are separate facts, and retrieval results do not change those states.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together; the model identifier should be checked first and non-matching bases excluded. Cross-model retrieval is not performed to compare missed recalls. Search modes include semantic search, full-text search, and hybrid search. Hybrid search merges multiple result paths through RRF, which is a fusion process and not a standalone switch. The minimum relevance score takes effect only under semantic search or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls the maximum content cited per search. Question optimization performs coreference resolution and question expansion, adding one model call. Image retrieval requires corresponding image input and model capability, and reranking requires a ReRank model enabled by configuration; these optional paths do not mean every request executes all of them. An empty result is output as an array whose length can be 0.
Hypothetical drill sample: the original text in a Tmall flagship store internal training material reads: “Brand authorization letter number ZS-2024-017, authorized party Hangzhou Trading Co., Ltd., authorization valid until 31 December 2025, scope covers flagship store sales of this brand on the Tmall platform.” A draft sentence to be assessed reads: “The store brand authorization expired in June 2024 and authorization materials must be resubmitted.” When checking, compare the draft sentence with the authorization validity field in the original. The original states validity until 31 December 2025; the expiry date in the draft sentence does not match the original and is unsupported.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Explanation |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases using the same vector model | Cross-model association is not allowed; check the identifier first and exclude non-matching bases |
| Search mode | Compare semantic, full-text, and hybrid search against test questions | No single mode is presumed suitable for all materials |
| Citation limit | Unit is tokens; test content completeness with materials of different lengths | For this hypothetical trial the starting value is 2000 tokens, chosen to cover most ticket paragraphs, compared with 1000 and 4000 |
| Minimum relevance score | Takes effect only under semantic search or when result reranking is used | Label relevant and irrelevant samples before adjusting the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Requires a ReRank model enabled by configuration | Rerank scores can be used for filtering; first-place hit is not guaranteed |
Three common errors
- Treating a retrieval result as authorization evidence. The input is the draft sentence “The store brand authorization expired in June 2024”; the comparison object is the original authorization validity until 31 December 2025; the observable result is that the two date fields do not match and the draft sentence is unsupported. Retrieving the original text does not mean the authorization is in effect or has been renewed.
- Treating a pending-review status as missing operation permission. The input is the handover request “Please confirm whether a refund can be executed for this after-sales ticket”; the comparison object is the ticket’s current status field; if the original only states “pending review,” the observable result is that approval completion is not recorded, and the ticket cannot be judged approved or rejected.
- Associating knowledge bases built on different vector models. The input is two knowledge base identifiers and their vector model fields; the comparison object is whether the model identifiers match; the observable result is that non-matching bases should be excluded, and cross-model retrieval is not performed to compare missed recalls.
How to verify the setup
- Input the original text and the draft sentence from the hypothetical drill sample, compare the authorization validity field, and if the draft sentence’s date does not match the original, the draft sentence is unsupported; if the original does not state that field, record it as pending verification.
- Input a test question containing an explicit number, run semantic search and full-text search separately, compare whether the recalled content contains the entry corresponding to that number, and observe whether full-text search more readily hits the exact number. Full recall is not guaranteed.
- Input an empty-citation scenario, check whether the downstream design accepts an array of length 0, and observe whether the output path completes. Do not claim that a fictional rule search necessarily returns an empty array.