What the documents look like
Product information and selection questions for a Tmall flagship store draw on four document types: product publishing pages, brand authorization letters, delivery commitment records and after-sales tickets. The authorization letter determines whether a brand may present goods under a flagship store identity. The publishing page carries category, specification and selection information. The delivery commitment record constrains the 48-hour dispatch wording. After-sales tickets show how buyers ask about models and accessories. These materials update daily and per order, so chunked indexes must keep pace with order volume.
Brand authorization, flagship store identity and Tmall points compensation rules carry more weight in this category than in a JD POP store. If an authorization letter and a publishing page are chunked poorly, qualification conditions can be split apart, and selection answers may give a buyer a model that does not apply. Downstream consequences described for this category include product delisting, delayed dispatch compensation and a drop in store rating.
Constraints at the document chunking step
Chunking must keep a condition, its explanation and the content it constrains inside the same segment. If the validity period, authorized brand and authorized store of a brand authorization letter are separated, a retrieved fragment cannot point back to the full authorization scope. If a specification table and its applicable models on a publishing page are cut into different chunks, selection answers can mistake an accessory for a main unit.
One set of source data can map to multiple vector indexes. During retrieval, if any one of those vectors is recalled, that data group is also recalled. This gives long authorization letters and specification tables room for annotation, but it requires chunking to preserve document identity and version that can point back to the original text. After data changes, a successful upload does not equal updated retrieval; actual training and index status need to be checked.
The hypothetical drill sample for this page is as follows. The source text for comparison reads: "Party A authorizes Party B to operate the 'Xingye' brand flagship store on the Tmall platform, with an authorization period from January 1, 2025 to December 31, 2025, and an authorization scope of Xingye brand rice cookers." One draft sentence to be judged reads: "Party B may continue to sell Xingye rice cookers under the Xingye flagship store name in 2026." The subject of this draft sentence is Party B, the object is Xingye rice cookers, the state is continued sales in 2026, and the effective range exceeds the authorization period.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | Selection questions contain both model numbers and natural language descriptions; compare hit differences against semantic retrieval alone |
| Reference limit | 3000 tokens as the hypothetical starting point for this drill | Authorization letters and specification tables have long paragraphs; compare against lower values to see whether conditional sentences are truncated |
| Minimum relevance | 0.4 as the hypothetical starting point for this drill | Compare against 0.2 and 0.6 to observe whether fragments containing the validity period are filtered out |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare against the disabled state for recall completeness on consecutive follow-ups |
| Result reranking | Enabled when candidates are numerous | Compare against no reranking for ranking stability on text questions |
| Index annotation | Multiple index sets annotated for the same authorization letter | Compare against a single index to observe whether period and brand are hit in the same chunk |
Three common errors
- Splitting the authorization period and the authorized brand of a brand authorization letter into different chunks. The input is the authorization letter text above. Compare whether the period sentence and the brand sentence fall in the same data group. If retrieval recalls only the brand sentence and misses the December 31, 2025 period, a selection answer may give a reply beyond the period.
- Treating the draft sentence to be judged as an already effective conclusion. The input is "Party B may continue to sell Xingye rice cookers under the Xingye flagship store name in 2026." Compare the year field of the draft sentence against the authorization letter text. The year 2026 falls outside January 1, 2025 to December 31, 2025, so the draft sentence is not supported.
- Treating a successful upload as updated retrieval. The input is the updated authorization letter file. Compare the upload record against training and index status. If the index still points to the old version, retrieval returns the expired period.
How to verify the setup
- Use "December 31, 2025" from the authorization letter text for full-text retrieval. Input that date string and compare whether the recalled fragment contains the authorized brand and authorized store. If all three return in the same chunk, chunking has preserved the constraint relationship.
- Use the draft sentence for semantic retrieval. Input that sentence and compare the recalled content against the year field in the source text. If the returned fragment clearly shows authorization ending on December 31, 2025, the system has not treated the draft sentence as an effective fact.
- Use an empty result to check the downstream design. Input the explicit paper input [] and compare whether the downstream handles it as no reference content. If the downstream does not treat the empty result as authorization approval, retrieval has not overstepped to replace business approval.
The above are paper checks. No actual retrieval, training, approval or compensation operation was executed.