What the documents look like
Daily operations for a Tmall flagship store center on the product publishing page, the brand authorization letter, the delivery commitment record, and after-sales tickets. Compared with a JD POP store, a Tmall flagship store places more weight on brand authorization, flagship identity, and Tmall points compensation rules. Materials come from the Tmall merchant backend, the Tmall rule center, ERP order tables, and customer service chat logs, with updates recorded daily and per order.
For product information and selection questions, this mix means the retrieval step handles both natural-language phrasing and precise lookups such as item numbers, model designations, and proper nouns. Vector similarity alone does not decide whether a draft sentence is supported; the retrieved original text has to be checked against the claim.
Constraints at the vector indexes step
FastGPT builds knowledge bases with an Embedding approach, where a smaller vector distance indicates greater similarity. Factors affecting vector search precision include vector model quality, data quality, and retriever precision. One data block can map to multiple vectors. When index content is more focused, retrieval precision rises while retrieval range narrows. During retrieval, vectors are recalled first, then the vector ID is used to find the original data in the MongoDB dataset.datas table. If several vectors point to the same original data block, the results merge and the highest vector score is kept. Knowledge bases linked for retrieval must use the same vector model. This page checks only the actual model identifier and input type; it does not extend that requirement to external systems. Image retrieval also depends on whether an image description or image vector index was generated at ingestion time.
Hypothetical paper exercise: A Tmall flagship store product publishing page draft states, "This store sells a certain brand of vacuum flask, capacity 500 ml, delivery commitment is dispatch within 48 hours after payment." The brand authorization letter states the authorized validity period runs from January 1, 2026 to December 31, 2026. An after-sales ticket records, "Buyer asks whether seven-day no-reason return is supported." One draft sentence to be assessed reads, "This product supports seven-day no-reason return, and the delivery commitment is 48 hours." The subject is a store operations staff member, the objects are the materials above, the state is draft and pending verification, and the effective scope follows the period stated in the authorization letter.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Search mode | Hybrid search | The sample contains both natural-language phrasing and precise terms such as capacity and model, compared against semantic-only results |
| Reference limit | Hypothetical starting point of 3000 tokens | Set because chunk lengths in the sample vary widely, compared against 2000 and 4000 for completeness of quoted original text |
| Minimum relevance | Hypothetical starting point of 0.4 | Compared against 0.3 and 0.5 for candidate count after filtering; effective only under semantic search or result reranking |
| Question optimization | Enabled | The sample contains follow-up turns, compared against the disabled state for the retrieval terms after coreference resolution |
| Result reranking | Enabled | Second-pass ordering when candidates are numerous, compared against RRF fusion alone |
| Vector model | Consistent with the linked knowledge base | Check the actual model identifier and input type first, exclude mismatches, and do not run cross-model retrieval |
Three common errors
- Treating a vector score as a conclusion about authorization or status. In the sample, the authorization letter's validity period and the draft sentence's return commitment are different fields. A high vector similarity does not mean the draft sentence is supported. The check takes the draft sentence and the authorization letter text as input, compares whether they share the same subject, the same period, and the same field, and observes that the draft sentence is not supported.
- Writing a pending-review state as approved. In the sample, the product publishing page is a draft and the brand authorization letter is pending verification. Neither constitutes a basis for listing or compensation. The check takes the draft state marker and the draft sentence as input, compares whether a signed or effective record exists, and observes that the evidence gap remains.
- Treating retrieval as permission filtering. Linked knowledge bases must share the same vector model, and retrieval returns reference content only; it does not decide whether an operations staff member may view a given record. The check takes the linked knowledge base identifier and the access scope as input, compares whether the cited source falls within the authorized scope, and observes that an out-of-scope source requires separate verification.
How to verify the setup
- Input the hypothetical sample's draft sentence, "This product supports seven-day no-reason return, and the delivery commitment is 48 hours." Compare it with the same field's value in the product publishing page draft and the brand authorization letter text. Observe whether the return commitment in the draft sentence has a corresponding statement in the original text. If the original text does not state it, record the result as unsupported.
- Input the after-sales ticket question from the sample. Compare the retrieval terms with question optimization enabled and disabled. Observe whether the coreference is completed into a full question containing subject and object. If the retrieval terms still lack a subject, record the result as pending optimization.
- Input the capacity and model precise terms from the sample. Compare the quoted original text returned by hybrid search and by semantic-only search. Observe whether the item number and model are matched. If they are not matched, record that full-text or hybrid search is needed, and check whether the ingestion index was generated.
The reference output of the knowledge base search node is an array and may have length 0. A search that returns no content still completes the output path. When a paper input of an empty array is used, the downstream design can be checked against that empty result; this does not claim that any particular rule-based search will return an empty array.