What the documents look like
Site selection and store opening at a chain headquarters take place before a store begins trading. The work centres on lease negotiation and engineering handover rather than on supervising operating outlets. Common inputs include trade-area footfall, map points of interest, historical store-type data from POS records, and landlord materials. The documents handled day to day are site assessment forms, letters of intent for lease, decoration budgets, and opening schedules. Materials are updated at irregular intervals, once per store, so retrieval usually faces several documents for the same store project at the same time.
These materials often differ in version and status. A letter of intent may sit at a pending-review stage, a decoration budget may exist only as a quotation, and an opening schedule may not yet be confirmed. The retrieval system returns text fragments and does not judge what authorisation status a document holds. The reference limit is measured in tokens, so materials with very different lengths affect how complete the returned content is.
For procurement and supplier records in this category, the practical consequence is that a retrieved fragment tells you what a document says, not whether it is signed, approved, or binding. A site assessment form, a letter of intent, a decoration budget, and an opening schedule can all be recalled together even though they carry different business standing.
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 must be checked first, non-matching bases excluded, and the actual observation recorded. Search modes include semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process rather than a switch that can be turned on. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission or entity isolation. Reference content is output as an array and its length can be 0, so the output path still runs when nothing is retrieved.
Hypothetical paper exercise sample: the source material reads, "Party A intends to lease a certain shop unit to Party B for chain-store operation. The letter of intent is dated 10 March 2025, the term is five years, and rent is paid quarterly. Party B shall submit a decoration budget to Party A for confirmation within fifteen days of signing. This letter of intent is a negotiation document and does not constitute a formal lease contract." The draft sentence to be assessed is: "The letter of intent has taken effect, and Party B may arrange decoration access on that basis." The source states the letter of intent is a negotiation document and does not constitute a formal lease contract, so the draft sentence's claim that it "has taken effect" is not supported by the source. The finding is unsupported, and the evidence gap is retained.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Store-expansion bases sharing one vector model | Check that vector model identifiers match before associating; exclude non-matching bases |
| Search mode | Hybrid retrieval | Run the same question under semantic and full-text retrieval separately, then compare recalled content |
| Reference limit | Hypothetical starting point of 800 tokens | Test with site assessment forms of different lengths to see whether content is truncated, then adjust |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare filtering results at different thresholds |
| Question optimisation | Enabled | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Enabled according to settings | Where a rerank model exists, compare ordering before and after reranking; rerank scores range from 0 to 1 |
The values above are hypothetical starting points for this exercise. They are not official defaults and need to be adjusted after comparison with actual materials.
Three common errors
- Treating retrieval recall as authorisation confirmation. Recalling a fragment of a letter of intent does not mean the letter has been signed or has taken effect. A pending-review status does not demonstrate that an operation permission is missing.
- Treating minimum relevance as a permission or entity filter. That value only filters low-relevance results. It does not exclude unrelated businesses or unauthorised records.
- Treating RRF as a switch that can be turned on. RRF is the fusion process for multiple result paths in hybrid retrieval, not a standalone configuration item.
How to verify the setup
- Enter the letter of intent source text from the hypothetical sample together with the draft sentence. Compare the draft sentence's "has taken effect" against the source wording "does not constitute a formal lease contract". Observe whether the system returns the source fragment and refrains from treating the draft sentence as an established fact. The condition is met when the returned content contains the source sentence and the draft sentence is not treated as confirmed.
- Enter the same question under semantic retrieval and full-text retrieval separately. Compare the recalled content from both paths and observe whether the reference content after RRF fusion in hybrid retrieval contains fragments hit by both paths. The condition is met when the fused result covers both sources. First-position hits are not guaranteed.
- Enter an empty-reference scenario using an explicit paper input of []. Check the downstream design and observe whether the output path still runs. The condition is met when downstream nodes can still execute with a reference array of length 0. No claim is made that a fabricated rule search will necessarily return [].
Across both verification steps and the configuration table, the same constraints apply. A retrieval library requirement for one shared vector model does not extend to business registration sites or other external systems needing the same model. Cross-model knowledge bases cannot be associated together; check the identifier, exclude non-matching bases, and record the actual observation rather than running a cross-model search to compare missed recalls. Optional paths in the documented flow, such as image retrieval, require a corresponding image input and model capability, and reranking requires an available model enabled according to settings. A pending review does not demonstrate that an operation permission is missing. Signing, reaching cooperation, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Where the source is silent or shows only a related stage, the finding is unsupported or pending verification, and the evidence gap is retained.