What the documents look like
Inventory procurement and supplier records for a Tmall flagship store arrive from the Tmall merchant backend, ERP order tables, and customer-service chat logs, together with brand authorization letters, dispatch-commitment records, and after-sales work orders. Compared with a JD POP store, a Tmall flagship store places more weight on brand authorization and flagship identity, and record errors can lead to product delisting, delayed-shipment compensation, and a lower store rating. This material appears as product publication pages, scanned authorization letters, dispatch-commitment text, and after-sales work orders, and it updates daily or per order.
At ingestion time, a longer text can be mapped to several vector groups, and one group of source data can correspond to multiple indexes. A retrieval hit on any one index recalls that group of data. The smallest search unit is the library; a collection only classifies data and does not affect search results. A similarity score does not carry permission or entity-isolation duties, and a retrieval hit does not constitute authorization confirmation.
Constraints at the document chunking step
Authorization letters, dispatch commitments, and after-sales work orders each contain a subject, an object, a validity scope, and a status. Chunking must keep a condition in the same segment as the content it constrains, so that an authorization period is not separated from the authorization object. Source data and retrieval indexes must be distinguished: one group of source data can map to multiple indexes, and an index is a retrieval entry point rather than a business authorization.
Semantic retrieval relies on vector distance. Full-text retrieval suits numbers, models, and proper nouns. Hybrid retrieval fuses multiple result paths through RRF, and reranking runs when configured and requires model support. The search scope is the whole library. A similarity threshold filters low-relevance results and applies only under semantic retrieval or reranking; it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Hypothetical paper exercise sample: the source text of a brand authorization letter reads, "Party A authorizes Party B to sell A-series products in the Tmall flagship store; the authorization period runs from January 1, 2024 to December 31, 2024; the authorization scope is limited to mainland China." A proposed draft sentence reads, "Party B may continue selling A-series products in 2025." Checking compares the authorization-period field. The draft sentence exceeds the end date stated in the source text, so it is unsupported.
Configuration and implementation checks
| Configuration or check item | Relevant setting on this page | Implementation note |
|---|---|---|
| Search mode | Semantic retrieval, full-text retrieval, hybrid retrieval | Choose by material type; number-type queries favor full-text or hybrid |
| Result reranking | ReRank model | Requires model support and runs when configured; the rerank score is 0-1 |
| Reference cap | Each search references at most n tokens | The value is adjusted against actual results; no default is specified here |
| Minimum relevance | A 0-1 value | Applies only under semantic retrieval or reranking; it is a filter threshold |
| Question optimization | Coreference resolution and question expansion | Adds one model call before retrieval; enable according to the scenario |
| Index check | Source data and the indexes field | Confirm whether a valid index was generated at ingestion; a successful upload does not mean retrieval is updated |
Three common errors
- Treating the authorization period in the letter and the sales time in the draft sentence as the same field. Input: the authorization letter text above and the draft sentence. Comparison: the authorization-period field. The draft sentence exceeds December 31, 2024, so the result is unsupported; it is not a contradiction and not a negation.
- Treating a retrieval hit as authorization confirmation. Input: one authorization-letter fragment returned by a semantic search. Comparison: whether that fragment states the authorization object and scope. The retrieval result only shows the content was recalled; authorization status requires a separate check of the source text.
- Treating a successful upload as an updated index. Input: the upload record after a material change. Comparison: the actual training and index state. When training completion is unconfirmed, whether retrieval reflects the new content remains unknown.
How to verify the setup
- Use the hypothetical sample above to check chunk completeness. Input: the three sentences of the authorization letter. Check whether the authorization period, authorization object, and scope stay in the same segment. If a condition and its constraint are split apart, the chunking does not meet the requirement.
- Search separately with a term, a number, and a natural-language question. Input: "A-series," "December 31, 2024," and one natural-language question. Observe whether the returned fragments contain the corresponding source text. When the result is empty or a fragment is missing, record it as pending verification and do not infer a cause.
- Recheck training and index state after a material change. Input: the changed material and the upload record. Compare the actual training state with the indexes field. When training is incomplete, whether retrieval is updated remains unknown. An empty result can be checked against the downstream design with the paper input []; this does not claim that a fictional rule search always returns [].