What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand. Store managers and regional managers sit inside one organisation and share service and training duties, so no franchise contract sits between the brand and the store. The material handled day to day includes daily sales reports, shift schedules, display photographs and inventory transfer forms. These materials are updated daily or weekly, and retrieval needs cluster around shift handover, checking training wording and supporting a transfer request.
The absence of a franchise contract matters for retrieval. The subject, store identifier, shift and authorisation state written in each document must remain visible in the returned reference content, because the reader has no intermediate contract layer to consult. Display photographs are image material. Whether they can be retrieved depends on whether an image description index or an image vector index was created when the material was stored, and on whether the current vector model accepts image input. Shift schedules and transfer forms carry approval states. Retrieval returns reference content only; approval and execution belong to downstream systems.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be associated together. Search modes are semantic search, full-text search and hybrid search. Hybrid search merges multiple result paths through RRF; RRF is a fusion process inside hybrid search, not a separate switch that can be turned on. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. Question optimisation performs coreference removal and question expansion and adds one model call before retrieval. Image retrieval requires image input and the corresponding model capability. Reranking requires a rerank model and must be enabled in the settings. The optional paths listed in the documentation do not all execute on every request. Reference content is output as an array, and its length can be zero.
A hypothetical paper exercise sample follows. The material text supplied for checking consists of three sentences. "Store A, 10 March 2025, day shift schedule is morning shift sales assistants Li Min and Wang Qiang, shift 09:00 to 18:00, store manager Zhang Hua approval state: pending review." "Store A, 10 March 2025, display photographs, six images, capture time 14:20, captured by Wang Qiang, no image description attached." "Store A, 10 March 2025, inventory transfer form number DB-0310-02, requested outgoing item spring knitwear size M, quantity 5, applicant Li Min, regional manager approval state: pending review." The draft sentence to be judged is: "The day shift schedule for Store A on 10 March 2025 has been approved by the store manager, the transfer request has been approved by the regional manager, and the outgoing transfer can be executed."
Configuration and implementation checks
The table below lists settings already described in the official documentation, or implementation checks, in relation to the sample above. The numeric values are starting points for this hypothetical exercise only, used for comparison against the actual material in the sample. They are not default values.
| Configuration item | Hypothetical exercise starting point and reason | Comparison and checking method |
|---|---|---|
| Associated knowledge bases | Select knowledge bases holding store service and internal training material, with matching vector model identifiers | Check the knowledge base identifier against the vector model identifier, exclude any that do not match, and do not run retrieval across different models |
| Search mode | Start with hybrid search, covering both natural language questions and identifier queries | Ask the same question and observe the reference content returned by the semantic, full-text and hybrid paths separately |
| Reference limit | Hypothetical starting point of 800 tokens, because the sample material is short and the first aim is to see whether all three sentences return | Compare against the length of the sample text and check whether the schedule, photograph and transfer sections are truncated |
| Minimum relevance | Hypothetical starting point of 0.3, to be adjusted after marking relevant and irrelevant passages in the sample | Compare the reference array before and after filtering, only under the conditions where semantic search or result reranking is in effect |
| Question optimisation | Start with it off, because the sample question already names its subject clearly | Compare the original question with the optimised question and check whether the store identifier, date or shift has been altered |
| Result reranking | Start with it not enabled, because the sample has few candidates | Compare candidate ordering only after a rerank model is available and enabled in the settings; when it is not enabled, record that it was not enabled |
Three common errors
- Treating "pending review" as already approved. In the sample, both the schedule approval state and the transfer approval state are pending review, while the draft sentence states that approval has been granted and the outgoing transfer can be executed. This writes an unfinished state as a completed one. The check compares the value of the same field for the same store and the same date; pending review and approved are mutually exclusive values of that field.
- Treating a retrieval hit as authorised execution. Returning a schedule or transfer form as reference content shows only that the material was recalled. It does not show that approval passed or that the outgoing transfer may proceed. The check separates behaviour the node already performs, configurable fields, manual checking actions, and business actions that require a separate downstream system.
- Forcing knowledge bases with different vector models into one association. The vector model identifiers of the sample knowledge bases must be checked first. Knowledge bases whose identifiers do not match are excluded and the actual observation is recorded. No inference about missed recall or redundancy is drawn from this.
How to verify the setup
- Enter the three sample sentences and the question "What are the approval states of the day shift schedule and the transfer for Store A on 10 March 2025". Compare whether the returned reference content contains the original wording "pending review", and check whether the draft sentence has been rewritten as approved. If the reference content is missing the approval state field, record the result as unsupported.
- Enter the question "Who captured the display photographs for Store A on 10 March 2025". Compare whether the returned reference points to Wang Qiang and 14:20, and check whether any indication of a missing image description index appears. If an empty array is returned, check the downstream design against the documented condition that output length can be zero; do not claim that a rule-based retrieval must return an empty array.
- Enter the question "Applicant and approval state for transfer form DB-0310-02". Compare whether the reference content from full-text search and from hybrid search hits the form number and "pending review", and check that subject, store, date and field agree. If the hit refers to another store or another date, record it as a mismatch in the same field and keep the evidence gap open.
This page covers only brand-owned store operations, store service and internal training, and knowledge base retrieval and recall. During operational handover, which materials, actions and requests sit at which authorisation state must be read from the documents themselves; retrieval and forms are not authorisation to execute.