What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand. They are not managed through franchise contracts, so procurement and supplier records sit alongside the same store-level materials that operations already produces: daily sales reports, shift schedules, display photographs and stock transfer orders. The brief describes these materials as updating daily or per order.
For procurement and supplier records, the documents that reach a knowledge base include quotation sheets, supply instructions and reconciliation attachments. Their subject, object, amount basis and validity scope need to be read against the store's actual receiving and dispatch records. A supply instruction may name a supplier, a category of goods, a date range and a condition attached to a transfer. When chunking splits a condition away from the content it constrains, a retrieved fragment can no longer be traced back to the original text.
The relevant distinction for this category is that brand-owned store staff are directly employed and measured by the brand. Procurement documents therefore connect to store sales and transfer data rather than to a franchise supervision relationship. The eight-dimension brief is a category description, not a record of any specific company, and its confidence level is moderate.
Constraints at the document chunking step
Chunking follows the structure of the material itself. Conditions, explanations and the content they constrain stay in the same segment. The official knowledge base documentation separates original data from retrieval indexes: one group of original data can map to multiple vectors, and the indexes field records the corresponding vector IDs. This multi-vector mapping exists because content length and semantic richness are usually in tension, and a longer passage can be given several vectors so that a match on any one of them recalls the data.
The reference limit is measured in tokens, not in a top-k item count. The documentation explains that token-based control was adopted because chunk lengths in mixed knowledge bases vary widely. Minimum relevance is a 0-1 value that filters low-relevance results, and it takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold, not a final ordering rule.
Hybrid retrieval combines semantic and full-text recall through RRF. Reranking requires a model and must be enabled according to settings. Image retrieval requires image input and the relevant model capability; with an ordinary vector model and no vision model, pure image queries are essentially unavailable. Knowledge bases built on different vector models cannot be associated with each other. The correct handling is to check identifiers and exclude those that do not match.
A hypothetical paper exercise sample is used below. It is clearly labelled and is not a record of any real supplier, store or approval.
Hypothetical exercise sample. A regional operations file for a brand-owned store retains the original text of a supplier supply instruction: "This instruction applies to suppliers supplying Category A knitwear to brand-owned stores between 1 March 2026 and 31 March 2026. Where a single transfer request exceeds 50 items, the regional manager must sign the transfer order to confirm." In the same batch of materials there is a draft sentence awaiting assessment: "This supplier does not require regional manager signature for store supply during March." The comparison material is the same supplier, the same period and the same field of supply conditions. The draft sentence and the original text take mutually exclusive values on the condition "signature required above 50 items," so the draft sentence is unsupported.
Configuration and implementation checks
The table lists only settings or implementation checks that the official documentation describes and that relate to document parsing and chunking for this page. The hypothetical starting values are for this paper comparison only and are not official defaults.
| Configuration item | Hypothetical exercise starting value | Reason and comparison object |
|---|---|---|
| Search mode | Hybrid retrieval | Supply instructions contain both identifiers and natural-language conditions; compared with semantic retrieval alone for recall coverage |
| Reference limit | 3000 tokens | Intended to cover a full supply condition passage; chosen after comparing with the length of the original passage |
| Minimum relevance | 0.35 | Filters low-relevance fragments; compared with starting values of 0.2 and 0.5 |
| Question optimisation | Enabled | Completes references in follow-up questions about supply scope; compared with the disabled state |
| Result reranking | Enabled | Performs secondary ordering when candidates are numerous; compared with the disabled state |
| Image index | Checked against model capability | On display photograph ingestion, confirm whether an image description index or image vector index was generated |
Three common errors
- Treating "signature required above 50 items" and "no signature required" as two phrasings of the same condition. The input is the draft sentence and the original text; the comparison is the same supplier, the same period and the same field; the observable result is whether the values are mutually exclusive. Mutual exclusivity means the draft sentence is unsupported. If the draft sentence were changed to "no signature required within 50 items," it would not be mutually exclusive with the original text.
- Treating a successful upload as proof that retrieval has been updated. The input is the revised supply instruction; the comparison is the actual training and index state; the observable result is whether retrieval returns the new condition. A miss leaves the update pending review, and the actual training state remains unknown unless it is observed.
- Using minimum relevance to exclude unrelated businesses or unauthorised records. The input is a similarity threshold; the comparison is the returned content against the permission scope; the observable result is whether out-of-scope fragments are still returned. A relevance score is not a permission filter, and retrieval is not authorisation.
How to verify the setup
- Enter the hypothetical sample original text and the draft sentence. Compare the values of "signature required above 50 items" and "no signature required." Observe whether the result is assessed as unsupported. If the draft sentence is changed to "no signature required within 50 items," it is not mutually exclusive with the original text. This is an expected outcome for the paper exercise, not a completed test.
- Enter two kinds of retrieval terms, such as "Category A knitwear transfer signature" and "March 2026 supply conditions." Compare whether the returned fragments contain the complete condition sentence. Observe whether the condition and the content it constrains remain in the same segment. A first-hit guarantee or a full-recall guarantee is not assumed.
- Enter a display photograph and check the ingestion index. Compare whether an image description index or an image vector index was generated. Observe whether a pure image query can match. Where no vision model is configured, record that this path is unavailable. The optional paths in the overall retrieval flow are not executed on every request.
Where a document field is not provided, it remains unknown. A pending request is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Missing evidence stays missing.