What the documents look like
Tmall flagship store operations rely on product listing pages, brand authorization letters, delivery commitment records, and after-sales tickets. These materials come from the Tmall merchant backend, the Tmall rule center, ERP order tables, and customer service chat logs, with updates occurring daily and per order. Flagship identity and brand authorization are the facts that separate this category from other store forms. An authorization letter carries a granting entity, a granted entity, and a validity scope. When parsing splits these fields apart, later comparison loses its anchor for checking.
The material formats differ widely. Product listing pages lean toward structured fields, authorization letters toward paragraph clauses, and tickets toward conversational records. Parsing and chunking need to follow each structure rather than forcing one shape onto all of them. Delivery commitment records and after-sales tickets are produced per order, so the delivery deadline, compensation basis, and ticket conclusion for the same order must stay aligned with the rule text.
Constraints at the document chunking step
In FastGPT, a knowledge base consists of three layers: library, collection, and data. The smallest search unit is the library; collections serve only to classify and manage data and do not affect search results. Vector retrieval uses the PostgresSQL PG Vector plugin with an HNSW index. The MongoDB dataset.datas table stores original data and includes an indexes field that records vector IDs, so one data group can map to multiple vectors. During retrieval, vectors are recalled first, then the original data is located in MongoDB by vector ID. When vectors correspond to the same original data group, they are merged and the highest vector score is taken.
These conditions mean chunking must preserve the document identity and version that allow a return to the source text. Conditions, explanations, and the content they constrain must not be separated.
Hypothetical drill sample: A brand e-commerce department submits a scanned brand authorization letter on March 10, 2026. The source text contains three sentences — "Granting party: Jia Brand Management Co., Ltd." / "Granted party: Yi E-Commerce Co., Ltd." / "Authorization period: January 1, 2026 to December 31, 2026; authorization scope: Tmall flagship store sales." The same-day handover request reads: "Please confirm this authorization letter can be used directly for this store's product listing page qualification submission." The request is in a pending review state and does not constitute evidence that authorization has taken effect or that platform approval has passed.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason for choice and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | The authorization letter contains entity names and dates, so both semantic and keyword coverage are needed; compared with pure semantic retrieval |
| Reference limit | Hypothetical starting point of 3000 tokens | The authorization letter paragraphs are short; compared with behavior when no value is set, to confirm entity fields are not truncated |
| Minimum relevance | Hypothetical starting point of 0.3 | Compared with 0.1, to observe whether entity names are filtered out |
| Question optimization | Enabled | The handover request contains a reference; compared with the disabled state for recall in follow-up turns |
| Chunking basis | Segment by authorization letter clause | Compared with fixed-length splitting, to check whether granting party, granted party, and period stay in the same chunk |
| Index check | Inspect the indexes field | Confirm that one original data group maps to multiple vectors; compared with a single-index case for recall paths |
Three common errors
- Treating a successful upload of the authorization letter scan as retrieval already updated. Upload completes data writing only; training and index status require separate checking. When the indexes field in dataset.datas is empty, that data cannot be recalled by vector search.
- Treating minimum relevance as a permission or entity filter. That value is a 0–1 similarity threshold that filters low-relevance results only. It cannot exclude unauthorized records or materials from another entity.
- Treating the pending handover request as authorization already effective. The request is in a pending review state. Whether the authorization period and scope cover this store is a separate fact and cannot be inferred from the request itself as grounds for qualification submission.
How to verify the setup
- Input the three sentences of the hypothetical authorization letter and the request "Please confirm this authorization letter can be used directly for this store's product listing page qualification submission." Compare whether the retrieved content contains the granting party, granted party, authorization period, and scope together. If the returned fragment lacks the granted party or the period, chunking has separated entity fields and the segmentation basis needs adjustment.
- Input "Jia Brand Management Co., Ltd. 2026 authorization period" and compare hits using full-text retrieval and semantic retrieval separately. If full-text retrieval hits and semantic retrieval does not, record this as the keyword path being effective. If neither hits, check whether the data has completed training and whether indexes were generated.
- Input an empty reference limit scenario, using the paper input [] to check downstream reference trimming logic, and observe whether empty reference content is returned. This result describes only the handling of empty references under this input and does not indicate that any fabricated rule retrieval necessarily returns empty.
The above checks are paper-based reasoning and were not actually executed. They do not constitute a test, deployment, or approval conclusion.