What the documents look like
Site selection and store opening take place before a store begins trading. The work centres on lease negotiation and engineering delivery, which separates this category from post-opening supervision. The main inputs are trade-area footfall, map POI data, POS history by store type, and landlord materials. Day-to-day documents cluster into four kinds: site evaluation forms, letters of intent for lease, fit-out budgets, and opening schedules. Together these materials decide whether a store investment continues, and errors show up as store losses, delayed delivery, or deposit forfeiture. Materials update irregularly, per store, so retrieval faces multiple versions left by the same store at different stages of progress.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the base; a collection only classifies and manages data and does not affect search results. Semantic search relies on vector distance, full-text search relies on keyword hits, hybrid search merges the two paths through RRF, and reranking performs a second ordering of candidate text and returns a score from 0 to 1. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Question optimisation performs coreference resolution and question expansion before retrieval and adds one model call.
Retrieval itself does not judge permissions. A similarity score is not a permission filter, and retrieval results do not constitute proof of authorisation, approval, or completed handover. Citation content is output as an array, and its length can be 0; when nothing is found, the output path still runs through.
Hypothetical paper exercise sample: a chain headquarters development staff member handles a handover of one letter of intent for lease. The source text reads: "Party A is a commercial management company, Party B is a chain headquarters. The letter states that the leased subject is a third-floor unit in Mall A, that the intent deposit is payable within five working days from the date of signing, and that this letter takes effect from the date both parties sign, with a scope limited to negotiation of lease terms and not constituting a formal lease contract." The sentence to be assessed reads: "This letter of intent has taken effect, and Party B has obtained authorisation to enter and begin fit-out."
Configuration and implementation checks
The settings below relate only to this feature and to operations permission handover around the sample. The values are starting points for this hypothetical exercise, used for paper comparison, and are not official default values.
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Associated knowledge bases | Site evaluation form base, letter of intent base | Both bases must use the same vector model; check the identifier before associating |
| Search mode | Hybrid search | Compare with pure semantic search to see whether identifier-type terms are hit |
| Citation limit | 1200 tokens | Compare with 800 and 1600 to see whether letter clauses are truncated |
| Minimum relevance | 0.35 | Compare with 0.2 and 0.5 to see whether unrelated unit materials are brought in |
| Question optimisation | Enabled | Compare with disabled, and check whether the subject of "this letter" is altered incorrectly |
| Result reranking | Enabled | Compare with disabled, and observe whether candidate clause ordering changes |
Three common errors
- Treating a retrieval hit as authorisation. The input sentence to be assessed retrieves the source letter, and the output is citation content only. The comparison object is whether the effective scope of the letter and "authorisation to enter and begin fit-out" belong to the same field and the same subject. These are not the same fact, and the conclusion is recorded as unsupported.
- Writing a pending review state as a lack of permission. In the sample the letter sits at the negotiation stage, and the material does not state an approval step. Compare whether an approval field appears in the material. When it is absent, record the evidence gap rather than judging it as a denial or a conflict.
- Forcing association of bases with different vector models. Input the two base identifiers, check the vector model identifier first, exclude those that do not match, and record the exclusion. Do not run a cross-model retrieval to compare missed recalls.
How to verify the setup
- Input the sample source text and the sentence to be assessed, then compare whether the effective-scope field of the letter and the authorisation described in the sentence share the same subject, the same period, and the same field. Record a contradiction only when the same field carries clearly mutually exclusive values; when the field is not stated, record it as pending verification.
- Input a unit material that contains no authorisation clause, then compare citation content before and after the minimum relevance filter. Observe whether unrelated fragments are brought in. The failure branch is that unrelated fragments still appear in the citations; at that point record the actual observation and do not infer a system error cause.
- Input the explicit paper input
[], then compare how downstream nodes are designed to handle an empty array. Observe whether the output path runs through, and do not claim that a fictional rule retrieval necessarily returns[]. Business actions such as signing, contract effectiveness, and entry permits must be implemented by downstream systems; retrieval and forms do not carry that judgement.