What the documents look like
Tmall flagship store product information and selection Q&A draws on four material types: product listing pages, brand authorization letters, delivery commitment records, and after-sales tickets. Brand authorization and flagship store identity are the key facts that distinguish this category from other store formats. Whether a product can be listed and under what name it is sold depends on whether the authorization chain is complete.
Delivery commitment records are generated per order. Delayed delivery triggers compensation and affects store ratings, so any answer in selection Q&A that touches delivery timing must be traceable to a specific commitment record. After-sales tickets record the handling status of return, exchange, and repair requests. These materials together form the retrieval corpus, but their subject, period, and status fields are not uniform.
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 with different vector models cannot be associated together. During implementation, verify the vector model identifier first, exclude those that do not match, and record the actual observation. Do not execute cross-model retrieval to compare missed recalls.
Retrieval itself does not constitute authorization. Recalling a passage of brand authorization text only indicates that the text is similar to the question. It does not indicate that the current store has obtained authorization, that the authorization is within its valid period, or that the product can be listed.
The minimum relevance score takes effect only during semantic retrieval or when result reranking is used. It is a filtering threshold, not a permission or entity isolation mechanism. It cannot be used to exclude unrelated businesses or unauthorized records.
The reference limit is measured in tokens. Materials with large length differences need separate verification of whether content is complete. Question optimization is used for coreference resolution and question expansion, and it makes an additional model call. Compare the original question with the optimized question to verify whether the subject has been altered incorrectly.
Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion process and is not described as a switch. Image retrieval requires corresponding image input and model capability. Reranking requires a model and must be enabled according to settings. The optional paths listed in the overall flow diagram do not mean that every request executes all of them.
Configuration and implementation checks
The table below centers on the hypothetical drill sample for this page and lists officially documented settings or implementation check items. The values are starting points for this hypothetical trial, with the reason for selection and the comparison object stated alongside.
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base storing product information and selection Q&A corpus | Verify each base's vector model identifier first, exclude those that do not match, and record the actual association result |
| Search mode | Test semantic retrieval and full-text retrieval first, then test hybrid retrieval | Use the same set of test questions to execute each mode separately and compare the correspondence between recalled content and the question |
| Reference limit | Hypothetical starting point set to accommodate one complete delivery commitment record | Test with actual materials of different lengths to see whether content is truncated, and adjust after comparing with the complete record length |
| Minimum relevance score | Hypothetical starting point takes a middle value | Annotate relevant and irrelevant samples first, then compare the retention of both sample types under different values |
| Question optimization | Enable first in multi-turn follow-up scenarios | Compare the original question with the optimized question to verify whether the subject and object have been altered |
| Result reranking | Enable first when candidates are numerous | Compare candidate order before and after reranking and verify the relationship between scores and question completeness |
Three common errors
- Treating a retrieval hit as authorization confirmation. Input a piece of brand authorization text and retrieve it. What is compared is text similarity, and the observable result is that reference content is returned. Whether the authorization is valid and whether the product can be listed are separate facts. The authorization document's subject, object, and valid period must be verified separately. The retrieval result does not provide that conclusion.
- Using the minimum relevance score to exclude unauthorized records. This value takes effect only during semantic retrieval or result reranking. What is compared is content relevance, and the observable result is that low-scoring content is filtered out. It does not distinguish business ownership or permissions. Raising the threshold only reduces returned items and cannot replace authorization verification.
- Treating RRF after hybrid retrieval as a switch that can be turned on. RRF is the process of fusing multiple result paths. What is compared is the ranking from each recall path, and the observable result is that the same content ranks higher when hit by multiple paths. At the configuration level, verify the search mode and reranking settings. Do not set up a switch named RRF.
How to verify the setup
The following verification uses the same hypothetical drill sample. Sample material text (hypothetical): A brand authorization letter states that the authorizing party is Company A, the authorized party is Company B, the authorization object is a certain model of product, the authorization period is from a certain date to a certain date, and the purpose is sales in a Tmall flagship store. Sentence to be judged (hypothetical): "Company B has obtained authorization and can sell this model of product in this store." Request to be judged (hypothetical): Operations submits a handover request asking to list this model of product and citing the above authorization letter as the basis.
- Verify whether retrieval can recall the authorization letter text. Input the product model and authorization subject from the sentence to be judged. Compare whether the returned reference content contains the original authorization letter text. The observable result is that the authorizing party, authorized party, authorization object, and period fields appear in the reference content. When not recalled, record the gap and do not infer that authorization does not exist.
- Verify whether the sentence to be judged exceeds the scope supported by the original text. Compare the sentence to be judged with the original text. The original text states the authorization period and purpose. The sentence to be judged does not state the period or purpose. The observable result is that the sentence lacks period and purpose qualifications, is unsupported, and requires supplementary verification.
- Verify the authorization boundary of the handover request. Compare the request with the original text. The original text only describes the authorization relationship and does not state that the listing operation has been approved. The observable result is that the request is in a pending review state. Pending review does not prove the absence of operational permission, nor does it constitute approval. The evidence gap must be preserved. For the empty-result case, a clearly paper-based input of an empty array can be used to verify downstream design. Do not claim that a fictional rule retrieval necessarily returns an empty array.