What the documents look like
Site selection and store opening happen before a store starts trading, and the work centres on lease negotiation and engineering handover. The main document set is the site evaluation form, the letter of intent for the lease, the renovation budget and the opening schedule. Common data sources are trade-area footfall, map points of interest, historical store-type data from the point-of-sale system, and landlord materials. These materials are organised per store and updated at irregular intervals rather than on a fixed cycle.
The status fields across this material set differ sharply. A letter of intent may be awaiting review, signed, or unsigned. A budget may still be a calculation draft. A schedule may not yet be confirmed. Retrieval has to recall content according to the subject, object, stage and validity scope stated in the source text. A retrieval hit is not an authorisation, and it does not mean performance has already taken place.
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 model identifier should be checked first and non-matching bases excluded. A single retrieval combines question optimisation, semantic retrieval, full-text retrieval or hybrid retrieval, image handling, RRF fusion, ReRank reordering and citation trimming according to the input content and search parameters. These optional paths do not all execute on every request. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic retrieval or when result reordering is used. Semantic recall and relevance scores do not perform permission or entity isolation, and retrieval results do not constitute authorisation.
Hypothetical paper exercise sample: the source text to be checked reads, "Party A and Party B signed a letter of intent for a shop unit in a certain trade area on 10 March 2025. The letter states that both parties still need to complete approval of the formal lease contract, and that no lease relationship arises before approval is granted." The draft sentence to be judged reads, "The lease contract for this unit took effect on 10 March 2025, so renovation access can be arranged." Comparing the two texts shows that signing a letter of intent and a contract taking effect are different facts. The draft sentence turns a pending-approval status into an effective status and derives an access arrangement from it. That position is not supported by the source text.
Configuration and implementation checks
| Configuration item | Relevant setting or check on this page | Relation to the hypothetical sample |
|---|---|---|
| Associated knowledge bases | Select one or more knowledge bases that use the same vector model | The site evaluation form, letter of intent, renovation budget and opening schedule may sit in separate bases; the vector model identifier must match |
| Search mode | Compare semantic, full-text and hybrid retrieval against the test question | The sample contains a date and status wording, so full-text and hybrid recall can be compared |
| Citation limit | Unit is tokens; test with real materials of different lengths to see whether content stays complete | The letter of intent has longer clauses, so observe whether the approval condition is truncated |
| Minimum relevance | Takes effect only under semantic retrieval or when result reordering is used | Label relevant and irrelevant samples first, then adjust the threshold; no value is preset |
| Question optimisation | Used for coreference removal and question expansion | In a multi-turn follow-up about "this unit", check whether the subject is altered |
| Result reordering | Reorders candidates when a ReRank model is available and enabled by configuration | With many candidates, observe whether the approval-condition sentence is still recalled |
For this paper exercise, a starting citation limit of a number of tokens may be set, on the reasoning that the value needs to cover the full wording of the approval clause in the letter of intent. The comparison method is to increase and decrease the value step by step and observe whether the clause is truncated. This value is only a starting point for this paper exercise.
Three common errors
- Treating a retrieval hit as an authorisation to act. The input is the letter of intent text above. If the draft concludes that renovation access can proceed, the comparison object is the sentence "no lease relationship arises before approval is granted", and the observable result is that the draft contradicts the status field in the source text.
- Writing a pending-review item as signed or already effective. The input is the signing date and the approval condition in the letter of intent. The comparison object is the draft sentence "the contract took effect". The observable result is that the same field carries mutually exclusive values.
- Associating knowledge bases that use different vector models. The input is the vector model identifier of two bases. The comparison object is whether the identifiers match. The observable result is that differing identifiers require exclusion; no cross-model retrieval is run to compare missed recalls.
How to verify the setup
- Input the letter of intent text above together with the draft sentence. Compare whether the draft turns a pending-approval status into an effective status. The condition is not met when a status-field contradiction appears.
- Input the letter of intent and the renovation budget for the same unit. Compare whether the subject, period and fields agree across the two materials. When the period or subject differs, record the result as unsupported rather than as a denial or a conflict.
- Input a paper empty-result array. Check whether the downstream design passes through when the citation content length can be zero. The condition is met when downstream can accept an empty citation. No claim is made that a fictional rule will always return an empty array.
Retrieval checks record observations and failure branches only. They do not guarantee a first hit, full recall, or that an unrelated query returns nothing. Signing, reaching cooperation and a contract taking effect are three different facts, and a quotation validity period does not indicate a supply period. Missing evidence stays a gap, and a pending review does not demonstrate a lack of operating permission.