What the documents look like
Site selection and store opening work happens before a store begins trading. The work centres on lease negotiation and construction handover, and the working papers are site evaluation forms, lease letters of intent, fit-out budgets and opening schedules. These documents are refreshed per store on an irregular cadence, so several versions and several signature states can sit inside one batch of material. The structures differ sharply. A site evaluation form carries calculation inputs such as trading-area footfall and map POI data. A lease letter of intent carries clauses and a validity period. A fit-out budget carries itemised amounts. An opening schedule carries milestone dates.
Retrieval and recall must place these differently structured documents into one knowledge base and return citations that keep the subject, the object, the status and the effective scope visible, so that store service and internal training staff can check them. The public brief marks this category profile as broadly indicative rather than confirmed, so the descriptions above are working assumptions about the material, not verified facts about any particular chain.
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 built on different vector models cannot be associated together; the vector model identifier should be checked first and non-matching bases excluded. Search mode is compared across semantic retrieval, full-text retrieval and hybrid retrieval against test questions; no single mode is assumed to suit every document type. The citation cap is measured in tokens and controls how much content one search may cite. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used, and it is a filter threshold rather than a permission or entity-isolation control. Question optimisation performs coreference resolution and question expansion, and adds one model call before retrieval. Hybrid retrieval merges multiple result paths through RRF. Image retrieval requires image input and the corresponding model capability. Reranking requires a model and must be enabled in settings. Retrieval by itself is not authorisation; a form or a retrieval result does not indicate approval, signature or contract effect.
Hypothetical paper exercise sample: a chain headquarters development staff member submits one paper set for checking. The source material reads: "Site evaluation form (reference A-07) records the object as a candidate unit in the Chengbei trading area, evaluation status pending regional development review, evaluation date March 2024." "Lease letter of intent (reference B-12) records the lessee as a trading company, the leased object as the same unit, intent status signed intent, valid until 30 April 2024." The draft sentence to be judged is: "The lease for this unit has received headquarters approval and fit-out may be handed over." That draft sentence asserts two facts, approval and handover. The material records no approval entry and no handover confirmation, so both remain unsupported.
Configuration and implementation checks
| Configuration or check item | Value or action on this page | Basis and explanation |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases using the same vector model | The node requires the same vector model; check the identifier and exclude non-matching bases |
| Search mode | Hypothetical trial starting point: hybrid retrieval | The material contains both references and natural-language description; compare semantic, full-text and hybrid results before deciding |
| Citation cap | Hypothetical trial starting point: 800 tokens | This is a hypothetical starting value for this exercise; compare short and long site evaluation and budget fragments for content completeness |
| Minimum relevance | Hypothetical trial starting point: 0.35 | Label relevant and irrelevant samples first, then compare whether filtered citations still carry subject and status |
| Question optimisation | Enable for multi-turn follow-ups and short questions | Used for coreference resolution and question expansion; compare the original and optimised question for subject distortion |
| Result reranking | Enable per settings when a rerank model is available | Rerank scores run 0 to 1 and support filtering; do not enable when no model is configured |
Three common errors
- Treating a retrieval hit as authorisation. Input the draft sentence "The lease for this unit has received headquarters approval and fit-out may be handed over" and compare it with the the source material reads the site evaluation form status is pending regional development review and the lease letter of intent status is signed intent. Neither records approval or handover. The observable result is that the approval and handover asserted in the draft sentence are unsupported, and the evidence gap should be recorded.
- Treating the letter of intent validity period as a supply or handover deadline. Input the source wording "valid until 30 April 2024" and compare it with "fit-out may be handed over" in the draft sentence. The observable result is that the material records no handover date, the validity period does not express a handover deadline, and the item is marked as requiring verification.
- Associating knowledge bases that use different vector models. Input two knowledge base identifiers and compare their vector model identifiers. The observable result is that a base is excluded when the identifiers do not match; cross-model retrieval comparison is not performed, and only the observed association scope is recorded.
How to verify the setup
- Input the hypothetical sample source material and the draft sentence, then compare the approval and handover asserted in the draft sentence against the pending-review and signed-intent statuses recorded in the material. The observable result is that the facts asserted in the draft sentence have no corresponding entry in the material, so they are judged unsupported and the evidence gap is retained.
- Input the object fields for the same unit in the site evaluation form and the lease letter of intent, then compare whether the subject and object in the two documents agree. The observable result is that when the objects agree they are recorded as the same unit, while the status fields are recorded separately as pending regional development review and signed intent, without merging into approval.
- Input two knowledge base identifiers and compare their vector model identifiers. The observable result is that matching identifiers permit joint association, while non-matching identifiers lead to exclusion and a record of the observed retrieval scope. When citation content is empty, check the downstream design against an array length of 0; do not claim that a hypothetical rule must return an empty array.