What the documents look like
Closure settlement for chain stores runs toward exit and asset recovery, the reverse direction from site selection and store opening. The working materials are POS inventory records, fixed-asset systems, lease contracts and financial ledgers. The documents handled day to day are closure approvals, inventory clearance sheets, lease termination agreements and asset count sheets. These materials are updated per store on an irregular schedule, so a retrieval step has to distinguish between different stages of the same store. Operating reporting and exception review require the quantities, amounts and status fields inside those documents to be recalled and checked item by item. A retrieval result on its own does not constitute an approval, a signature or a settlement.
Constraints at the retrieval step
The knowledge-base search node can associate one or more libraries that use the same vector model. Libraries built on different vector models cannot be associated together; the vector model identifier has to be checked first and non-matching libraries excluded, with the actual observation recorded. Cross-model retrieval is not run in order to compare missed recalls. Semantic search, full-text search and hybrid search each suit different material. Precise lookups of numbers and model designations suit full-text or hybrid search. Hybrid search merges multiple result paths through RRF. The citation limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reranking is used. Question optimisation performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires an image input and the corresponding model capability; reranking requires a rerank model and must be enabled in the settings. The paths drawn in the documented flow diagram do not mean every request executes all of them. Retrieval and relevance scores do not carry permission or entity isolation.
Hypothetical drill sample: a regional operations team for a chain brand submits a paper closure approval whose body states "Store number A-017, closure baseline date 31 March 2025, status is pending regional finance review." Attached in the same batch is an inventory clearance sheet whose body states "Store number A-017, clearance quantity 480 items, status: draft." A draft sentence is also put forward for assessment: "Store A-017 completed closure on 31 March 2025, and the 480 items cleared from inventory have been settled." Checking against the source text, the approval status is pending review and the clearance sheet status is draft; the completed and settled claims in the draft sentence are not supported by the source text.
Configuration and implementation checks
| Configuration item | Starting point for this hypothetical trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge libraries | Associate only the libraries holding closure approvals, clearance sheets, lease termination agreements and asset count sheets | Check first whether the vector model identifiers of the libraries match; exclude those that do not and do not associate them together |
| Search mode | Run semantic search first, then run hybrid search | Compare the recalled content of the two modes with the same test question; do not presume one mode suits all material |
| Citation limit | Starting point set to 2000 tokens | Test whether content is complete using real material of differing lengths, then adjust the limit |
| Minimum relevance | Starting point set to 0.5 | Label relevant and irrelevant samples first, then adjust the threshold; the value takes effect only under semantic search or result reranking |
| Question optimisation | Off first, enabled only for multi-turn follow-ups and short questions | Compare the original question with the optimised question and check whether the subject has been altered incorrectly |
| Result reranking | Not enabled first | Compare ordering before and after reranking once a rerank model is available and enabled in the settings |
These values are starting points for this hypothetical trial and are not official defaults.
Three common errors
- Treating a retrieval hit as authorisation. The input "A-017 closure approval" returns the approval text, whose status is pending regional finance review. If closure is treated as approved on that basis, the input and the comparison object are misaligned: the comparison is between the retrieval result and the authorisation status, and the observed text only states pending review. A pending review does not show that operating permission is absent, and it does not constitute approval.
- Reading a draft status as settled. The input "A-017 inventory clearance sheet" returns a clearance quantity of 480 items and a status of draft. If the draft sentence says "settled," the comparison is between the draft sentence and the clearance sheet status field, and the observed field value is draft; the draft sentence is not supported by the source text.
- Associating libraries across vector models. The input is two library identifiers, and they are associated together without checking whether the vector models match. The comparison is between the two vector model identifiers, and where the observed identifiers differ the non-matching library is excluded and the actual observation recorded, without inferring missed recalls or redundant results.
How to verify the setup
- Input the closure approval source text "Store number A-017, closure baseline date 31 March 2025, status is pending regional finance review," and compare it with the draft sentence "Store A-017 completed closure on 31 March 2025." Observe whether the completed status claimed in the draft sentence and the pending review status in the source text share the same subject, the same period and the same field. The condition is met when the draft sentence is changed to match the source status or is marked as unsupported.
- Input the inventory clearance sheet source text "Store number A-017, clearance quantity 480 items, status: draft," and compare it with the draft sentence "the 480 items cleared from inventory have been settled." Observe that the quantity field matches while the status field does not. The condition is met when the draft sentence no longer claims settlement and the draft status is retained.
- Input two knowledge library identifiers and compare their vector model identifiers. Where they match, the libraries may be associated together; where they do not, exclude the non-matching library and record the actual observation. For the empty-result path, a downstream design can be checked with the explicit paper input
[]; no claim is made that a fabricated rule search necessarily returns[].