What the documents look like
Closure settlement handles exit and asset recovery, running in the opposite direction from site selection and store opening. The material comes from POS inventory, fixed-asset systems, lease contracts and financial ledgers, and takes the form of closure approvals, inventory clearance forms, lease termination agreements and asset count sheets. These records are produced per store on an irregular schedule. Within one store, approval, clearance, termination and counting belong to different stages, and their field definitions and effective conditions are independent of each other.
For knowledge-base retrieval and recall, material is organised by library, collections serve only as classification management, and the smallest search unit is the library. Exit material for one store may be spread across several documents. A retrieved passage shows semantic proximity; it does not show which authorisation state that store currently occupies. Approval, inventory clearance, lease termination and asset counting are separate records, so a hit on one of them cannot stand in for the others.
Constraints at the retrieval step
Retrieval recalls candidates by vector distance, fuses and reranks them, and returns cited content as an array whose length may be 0; the downstream path still runs when nothing is found. Semantic search depends on model training quality and its precision is unstable. Full-text search suits numbers and proper nouns. Hybrid search fuses multiple recall paths through RRF. The minimum relevance value takes effect only under semantic search or when result reranking is used, and it is a filtering threshold; it does not perform permission or entity isolation. Question optimisation performs coreference resolution and question expansion and makes an additional model call.
Linked knowledge bases must use the same vector model, and bases with different models cannot be linked together. Verification therefore checks the identifier first and excludes those that do not match; it does not run cross-model retrieval to compare missed recalls. Image retrieval requires the corresponding image input and model capability, and reranking requires an available model enabled in the settings. The optional paths shown in the overall flow diagram do not mean that every request executes all of them.
Hypothetical paper exercise sample: the source material reads, "A store closure approval form states: applicant, operations specialist Li; reason for closure, lease expiring and not renewed; approval status, pending review by the regional operations lead; date of issue, 10 March 2025." "The store's asset count sheet states: count scope includes two refrigeration cabinets and twelve shelving units; counted by the store manager; review column blank." The draft sentence to be assessed is: "The store's closure approval has been granted, so inventory clearance handover may begin." Verification compares the approval status field recorded on the form with the status asserted in the draft sentence. Same subject, same period, same field, mutually exclusive values: this is a contradiction. The blank review column on the count sheet is unsupported and remains to be confirmed.
Configuration and implementation checks
| Configuration or check item | Assumed starting point for this exercise | Reason for the choice and comparison object |
|---|---|---|
| Linked knowledge bases | Link only one closure settlement library that uses the same vector model | Check whether the vector model identifiers match, exclude those that do not, then link |
| Search mode | Run semantic search and full-text search once each on the same test question | Compare whether cited content under each mode covers the approval status field text |
| Citation limit | Start from the token count needed to cover the body of one approval form | Unit is token; compare with actual material of different lengths to see whether content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust against those samples | Takes effect only under semantic search or result reranking; it is a filtering threshold |
| Question optimisation | Enable for follow-up questions containing references, disable for already explicit questions | Compare the original question with the optimised question and check whether the subject was altered |
| Result reranking | Compare candidate ordering after a rerank model is available and enabled in the settings | Rerank score is 0-1 and is used to compare candidate relevance |
Three common errors
- Treating a retrieval hit as authorisation. The input is the approval form text above together with the draft sentence. Judging that clearance handover may begin because text about "store closure approval" was retrieved ignores that the approval status is pending review. Pending review does not show an absence of operating authority, and it is not approval. The correct action is to compare the status field values and record the unsupported part.
- Forcing a link across model-different libraries. The input is two libraries with different vector model identifiers. Linking them together directly breaches the same-vector-model requirement. The correct action is to check the identifiers, exclude the one that does not match, and record the observation, without inferring what would happen after such a link.
- Treating an empty citation output as a business conclusion. The input is one retrieval that recalls nothing; the output citation array has length 0 and the path still runs. Concluding from this that the store has no closure material confuses a retrieval result with a business fact. The correct action is to record the empty result and separately check whether the material has completed training.
How to verify the setup
- Input the approval form text and the draft sentence from the hypothetical sample, and compare the value of the approval status field with the status asserted in the draft sentence. Mutually exclusive values in the same field constitute a contradiction, and the draft sentence should not serve as a handover basis. If the source text does not record that field, mark it as unsupported.
- Input the identifiers of two libraries whose vector model identifiers differ, and compare whether they match. A mismatch means the joint-link condition is not met. Exclude the one that does not match and record the actual observation; do not run cross-model retrieval to compare missed recalls.
- Input a paper retrieval and expect the citation array to be
[], then compare whether the downstream design still runs the path at length 0. If the downstream continues on an empty result and attaches no business determination, the design matches the documented behaviour. Whether a retrieval actually returns empty still depends on the retrieval itself and is not guaranteed.
This page covers only: chain-store closure settlement / operations authority and handover management / knowledge-base retrieval and recall. Which materials, operations and requests sit in which authorisation state during handover cannot be answered by treating retrieval or a form as an executed authorisation.