What the documents look like
Closure settlement handles exit and asset recovery, the reverse direction from head-office site selection and store opening. The main data sources are POS inventory, the fixed-asset system, lease contracts and financial ledgers. Day-to-day documents include the closure approval, inventory clearance form, lease termination agreement and asset count sheet. These materials are updated irregularly, per store, and store service and internal training need operations, finance, legal and the landlord contact to retrieve one shared version of the facts.
The document shapes differ in ways retrieval must respect. A lease termination agreement carries parties, object, signing and effective status. An inventory clearance form and an asset count sheet carry quantities and date ranges. A closure approval carries approval nodes and the boundary of authorization. Retrieval that cannot separate these fields cannot support training Q&A or handover checks.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases using different vector models cannot be associated together; check the identifiers first and exclude those that do not match. Search results are output as an array of references whose length can be 0, and an empty result still lets the output path run through. Semantic search, full-text search and hybrid search each fit different material: numbers, models and proper nouns suit full-text or hybrid search, while natural-language questions suit semantic or hybrid search. Hybrid search merges multiple retrieval paths through RRF, which is a fusion process. The minimum relevance value takes effect only under semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The reference cap is measured in tokens.
Hypothetical drill sample: the source material reads "Party A and Party B signed the Lease Termination Agreement for a certain store on 1 March 2025, agreeing to return the premises on 30 April 2025; as of 10 May 2025, no signature and seal page from either party appears on this agreement." The draft sentence to be judged reads "The store lease was terminated on 30 April 2025, and the deposit may be refunded as agreed." In this sample, signing, reaching agreement and contract effectiveness are three distinct facts. The draft sentence treats an agreement with no seal page as a completed termination, which the source does not support.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only one base that uses the same vector model | Check identifiers first and exclude bases with a different vector model |
| Search mode | Try semantic, full-text and hybrid separately on the sample question | Compare whether the three paths cover the agreement parties and date fields |
| Reference cap | Set to cover the length of one complete lease termination agreement | Test with actual materials of different lengths to see whether content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then adjust the threshold | Takes effect only under semantic search or when result reranking is used |
| Question optimization | Compare short questions and multi-turn follow-ups separately | Check whether the subject is altered after coreference resolution |
| Result reranking | Compare when a rerank model exists and is enabled in the settings | Rerank scores run from 0 to 1 and can be used for filtering |
Three common errors
- Treating a retrieved result as authorized execution. In the sample the agreement has no seal page, so retrieving that agreement does not mean termination is complete and does not mean the deposit is refundable. Retrieval and forms collect values; business verification, approval, database writes and external synchronization require separate downstream implementation.
- Using minimum relevance to exclude unrelated businesses or unauthorized records. That value is a filter threshold and does not perform permission or entity isolation; a similarity setting cannot replace an authorization judgment.
- Associating knowledge bases that use different vector models. Check identifiers first, exclude those that do not match and record the actual observation. Do not run a cross-model search to compare missed recalls.
How to verify the setup
- Enter the sample source material and the draft sentence, then compare whether the draft sentence contains wording such as "termination completed" that goes beyond the source. The source states only the signing and return dates and the absence of a seal page, so the draft sentence should be judged unsupported, and the missing seal page remains an evidence gap.
- Enter a handover request, "Please refund the deposit under this agreement," and compare whether the request carries an effective or authorized basis. With no seal page, the request stays pending verification and cannot be treated as approved or executed.
- Enter an explicit paper input of an empty result
[]and compare whether the downstream design continues through on an empty reference. Do not claim that a fictional rule will certainly return[]on retrieval; check only whether the empty-result handling matches the documented output whose length can be 0.