What the documents look like
Site selection and store opening work happens before a store begins trading, and its core materials concern lease and construction delivery. The main data sources named for this category are trade-area footfall, map POI data, POS history by store type, and landlord materials. The documents handled day to day include site evaluation forms, lease letters of intent, renovation budgets, and opening schedules. These materials are updated at irregular intervals and form one group per store.
For copy drafting, retrieval has to surface passages that can be cited in proposed marketing or merchandise wording: an area range for a store type, the stated basis for a rent-to-sales ratio calculation, or a description of a milestone in the opening schedule. The update rhythm of the source documents does not by itself constrain the retrieval feature described on this page.
A site evaluation form may state a target store type with an area range, a lease letter of intent may state a first-year rent-to-sales ceiling, and an opening schedule may fix a construction entry date. Draft copy that restates these values has to be checked against the wording actually held in the knowledge base, including any limits or qualifiers the original carries.
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, so identifiers are checked first and non-matching bases are excluded. Reference content is output as an array, and its length can be zero; when nothing is found, the downstream output path still runs.
Results are returned after similarity filtering and a reference ceiling trim, and the reference ceiling is measured in tokens. The minimum relevance value is a 0-1 number that filters out low-relevance results, and it takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not act as a permission or entity-isolation control. RRF is the fusion step that combines multiple retrieval paths in hybrid retrieval, and it is not described as a separate switch. Image retrieval requires image input and the corresponding model capability, and reranking requires a rerank model enabled in the settings; neither path is a necessary part of every request.
A hypothetical paper sample for checking: a chain headquarters development team holds a site evaluation form whose text reads: "The target store type is the community standard store, with an area range of 80 to 120 square meters. The lease letter of intent states a first-year rent-to-sales ceiling of 12 percent. The opening schedule sets the renovation entry milestone in March 2026." The proposed draft sentence to be judged is: "This store type does not exceed 120 square meters, the first-year rent-to-sales ratio is kept within 12 percent, and renovation starts in March 2026." The check compares whether the draft sentence matches the source on subject, object, state, and effective scope.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Site evaluation form base, lease letter of intent base | Both bases use the same vector model; check the identifier before associating |
| Search mode | Hybrid retrieval | The draft sentence contains an area figure and a percentage, so semantic and keyword recall are both wanted |
| Reference ceiling | Hypothetical starting point of 2000 tokens | Site evaluation form passages are long; compare content completeness against the shorter letter-of-intent clauses |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then observe whether source clauses survive filtering |
| Question optimization | Enabled | Compare whether the subject is altered during multi-turn follow-ups, against the original question |
| Result reranking | Enabled per settings | When candidates are numerous, re-sort and check the score against the source correspondence |
These values are hypothetical starting points for this trial and are not official defaults.
Three common errors
- Treating minimum relevance as a permission or entity filter. Input a query about another brand's site selection and compare whether returned content still includes material from this base; the relevance score does not carry the function of excluding unrelated businesses or unauthorized records.
- Associating knowledge bases that use different vector models. Input two bases with different vector model identifiers and compare whether the system allows both to be selected; check identifiers first and exclude non-matching bases, rather than running a cross-model retrieval to compare missed recalls.
- Describing RRF as a switch that can be turned on. Input a hybrid retrieval request and compare whether returned results went through multi-path fusion; RRF is a fusion step and no corresponding switch item exists in the configuration.
How to verify the setup
- Input the site evaluation form source text and the draft sentence from the hypothetical sample, and compare whether the draft's area wording matches the source range of 80 to 120 square meters. If the draft says "does not exceed 120 square meters," check whether the lower bound was dropped; when the condition holds, the reference content should include the full source range.
- Input the lease letter of intent source text and the draft sentence, and compare whether "first-year rent-to-sales ceiling of 12 percent" was rewritten as "kept within 12 percent." When the condition holds, the reference content should retain the "ceiling" and "first-year" qualifiers.
- Input a query with no matching material and compare whether the output reference array has length zero. A separate paper input of an empty array can be used to check the downstream design; this does not claim that any invented rule retrieval necessarily returns an empty array.