What the documents look like
Closure settlement handles exit and asset recovery, running in the opposite direction from head-office site selection and store opening. The working materials come from POS inventory, the fixed-asset system, lease contracts and financial ledgers. Typical documents include closure approvals, inventory clearance forms, lease termination agreements and asset count sheets. These records are updated irregularly, once per store, so retrieval at any moment faces several materials for the same store at a particular stage. Order fulfilment and after-sales answering has to state where a given lease exit, clearance batch or settlement currently stands, which material supports that position, and who handles it within what scope. The confidence note in the brief marks the category picture as broadly accurate, so the details below stay conditional where the brief is not specific.
Constraints at the retrieval step
The knowledge base search node can associate one or more libraries that use the same vector model; libraries with different vector models cannot be associated together. Retrieval runs against the whole library, and a collection only classifies data for management, with no effect on search results. Hybrid search combines semantic and full-text results through RRF, which is a fusion process rather than a switch that gets turned on. The minimum relevance value takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not carry permission or entity isolation, and retrieval by itself is not authorization. The reference cap is measured in tokens, so materials of different lengths are cut differently under one cap. Question optimization performs coreference resolution and question expansion, and can shift the subject, so each item needs checking. An empty reference is output as an array of length 0, and the downstream chain still runs.
Hypothetical paper exercise sample: one material states that "Party A signed the lease termination agreement on 10 March 2025, agreeing to complete clearance by 30 April; the inventory clearance form records submission on 20 March with status pending review." A draft sentence reads: "The store has completed the lease exit and settled all amounts." The comparison input is the source text above and the draft sentence, compared on subject, period and field. The source supports only signature and a submitted form awaiting review. It does not support "completed the lease exit" or "settled all amounts."
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Associated libraries | Only closure material libraries sharing one vector model | Different vector models cannot be associated; check the model identifier first and exclude non-matching libraries |
| Search mode | Run semantic search and full-text search once each | Compare hit differences between identifier-type questions and natural-language questions |
| Reference cap | Set to a token amount that holds the longest clearance form | Test with actual materials of different lengths and observe whether content is truncated |
| Minimum relevance | Annotate relevant and irrelevant samples before adjusting | Takes effect only in semantic search or reranking; compare results before and after filtering |
| Question optimization | Enabled for multi-turn follow-ups and short questions | Compare the original and optimized question and check whether the subject was altered |
| Result reranking | Enabled when candidates are numerous, with scores compared | Requires a rerank model and must be enabled in settings; it does not run on every request |
Three common errors
- Treating a retrieval hit as an authorization conclusion. The input is a clearance form whose status field reads pending review. Comparing that field, an output that states approved is an overreach beyond the material.
- Using the relevance threshold to exclude unrelated business. The input is two kinds of store question inside the same library. Comparing returns before and after filtering, the threshold does not perform entity or permission isolation.
- Merging signature, deal reached and contract effective into one fact. The input is the lease termination agreement text. Comparing whether effective conditions are stated, an agreement that states none can only be recorded as pending verification.
How to verify the setup
- Input the sample source text and the draft sentence given above, then compare whether the draft exceeds what the source supports. A draft containing "completed the lease exit" while the source holds only a signature and a pending-review form is judged unsupported.
- Input a clearance form number into full-text search and compare whether the returned content contains that number. When the return is empty, record the gap and do not infer that the rule must return an empty array.
- Input an attempt to associate a library with a different vector model and compare the model identifiers. A mismatch excludes that library; do not run a cross-model retrieval to compare missed recalls.
A paper input of an empty array can be used to check the downstream design that receives an empty reference, without claiming that any rule returns an empty array in practice. Optional paths in the documented flow, such as image retrieval or reranking, do not execute on every request; image retrieval needs image input and matching model capability, and reranking needs a model that is enabled in settings. Where a field is not stated, or only stage-related evidence exists, the result is unsupported or pending verification rather than a contradiction or denial. Pending review does not demonstrate a lack of operating authority. The materials needed for a check must sit inside the accessible target retrieval scope or in an explicit manual input; when they are absent, the gap is recorded as it stands.