What the documents look like
Closure settlement handles exit and asset recovery, and its direction runs opposite to headquarters site selection and store opening. The main data sources are POS inventory, the fixed-asset system, lease contracts and financial ledgers. The routine documents are closure approvals, inventory clearance forms, lease termination agreements and asset count sheets. These materials are produced per store and per batch, and updates are irregular rather than scheduled.
Product data and selection questions must locate material lines on inventory clearance forms, equipment models on asset count sheets and clause status in lease termination agreements. The material sits across multiple collections. Retrieval treats the library as the smallest search unit, while collections serve only as classification management and do not affect search results.
The practical consequence is that a question about a specific SKU, device model or clause state may need to draw on several document types at once, and the answer depends on which document actually carries the field in question. A quantity recorded on a clearance form and a status recorded on an approval are different fields from different sources, so retrieval output alone cannot settle whether a settlement step has been authorized or completed.
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 must be checked first, non-matching libraries excluded, and the actual observation recorded. This requirement applies to the associated knowledge libraries of this node. It does not extend to external systems such as business registration sites, which are not part of this association rule.
Search mode can be semantic retrieval, full-text retrieval or hybrid retrieval. The three modes should be compared against the test question rather than assuming one mode suits all closure materials. The reference cap is measured in tokens. Content completeness should be tested with materials of differing length before the cap is adjusted. Minimum relevance takes effect only in semantic retrieval or when result reranking is used. Relevant and irrelevant samples should be labelled before the threshold is compared.
Question optimization performs coreference resolution and question expansion. The original question and the optimized question should be compared to check whether the subject has been altered incorrectly. Hybrid retrieval merges multiple result paths through RRF, which is a result-merging process. Reranking requires a model and must be enabled according to configuration. Image retrieval requires a corresponding image input and model capability. The overall retrieval flow lists optional paths, and these paths do not all execute on every request.
Retrieval or form collection does not constitute authorized execution. During operations handover, the authorization state of materials, operations and requests must be checked separately. A pending review status is not an approval. An offered option is not execution. An unsigned termination is not a completed termination.
Configuration and implementation checks
Hypothetical drill sample: a chain brand operations staff member checks an inventory clearance form. The source text reads: "Store number A-017, clearance date 10 March 2025, item SKU-8842 Bluetooth earphones, quantity 12 units, status pending review." The draft sentence under review is: "The store's 12 Bluetooth earphones have completed clearance." The retrieval scope is limited to the closure settlement library, and associated libraries use the same vector model.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge library | Closure settlement library | Check that the vector model identifier matches before associating; exclude libraries on other models |
| Search mode | Hybrid retrieval | Compare semantic, full-text and hybrid modes on recall of the SKU number |
| Reference cap | Hypothetical starting point 2000 tokens | Test content completeness with long and short materials, compared against a shorter cap |
| Minimum relevance | Hypothetical starting point 0.3 | Label relevant and irrelevant samples, then compare thresholds |
| Question optimization | Enabled | Compare original and optimized questions, checking whether the subject is altered incorrectly |
| Result reranking | Enabled | Confirm the rerank model is available, then compare ordering before and after reranking |
These starting values are trial settings for this drill only. They are not official defaults, and the comparison method matters more than the number chosen.
Three common errors
- Treating a pending review status as completed clearance. The draft sentence states "have completed clearance," while the source status is "pending review." These are mutually exclusive values in the same field, so the draft is unsupported by the source text.
- Associating libraries that use different vector models. The node requires associated libraries to share the same vector model. The identifier must be checked and non-matching libraries excluded. Cross-model retrieval should not be performed to compare missed recall.
- Treating a retrieval hit as authorized execution. A returned reference does not mean clearance has been approved. Authorization state must be judged from the actual signing of the closure approval and the inventory clearance form.
How to verify the setup
- Input the hypothetical source text and the draft sentence "The store's 12 Bluetooth earphones have completed clearance." Compare the source status "pending review" with the draft "completed." The check condition is met if the retrieval result includes that clearance form and the draft is marked as unsupported.
- Input a query containing only the SKU number. Compare the recalled content from semantic retrieval and full-text retrieval. If full-text retrieval matches the number and semantic retrieval does not, record the difference between the two modes. No guarantee is made that the first result will match or that all items will be recalled.
- Input an empty reference array []. Check whether the downstream node executes. The check condition is met if the output path runs normally without error, confirming that empty-result handling is available. This is a paper check of downstream design, not a claim that a rule-based search returns an empty array.
For each check, the material must be within the accessible target retrieval scope or an explicit manual input. Where evidence is missing, record the gap rather than filling it. A signature, a reached cooperation and an effective contract are three distinct facts. A quotation validity period does not indicate a supply period. Missing or stage-only evidence remains unsupported or pending verification, and is not recorded as a contradiction.