What the documents look like
Closure settlement runs in the opposite direction from head-office site selection and store opening. It handles exit and asset recovery rather than acquisition and launch. The working materials come from POS inventory, the fixed-asset system, lease contracts and finance ledgers. The people involved include chain operations staff, finance, legal and the landlord contact. Records are produced per store on an irregular schedule. The day-to-day documents are closure approvals, inventory clearance sheets, lease termination agreements and asset count sheets.
For product and marketing copy drafting, retrieval has to locate wording inside those documents about clearance scope, asset status, lease term and settlement arrangements, then judge whether a proposed sentence has material support. The retrieval target is the whole library. A collection is only a classification layer and does not change search behaviour.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Libraries built on different vector models cannot be associated together. Check the model identifier first and exclude any library that does not match, rather than running a cross-model retrieval to compare what is missed.
Retrieval runs in semantic, full-text or hybrid mode. Hybrid retrieval merges the two result sets through the RRF formula. Result reranking uses a ReRank model for a second ordering pass. The minimum relevance value applies only under semantic retrieval or when result reranking is used, and it acts as a filter threshold. The citation cap is measured in tokens. Citation output is an array and its length can be zero, so the downstream path still runs when nothing is found.
A relevance threshold filters by similarity only. It does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Hypothetical paper exercise sample. The source material reads: "The closure approval for a given store has been granted. The inventory clearance sheet lists shelving and refrigeration units, fourteen items in total. The lease termination agreement states that the deposit is to be refunded within thirty days after handover." The draft sentence under review reads: "The store assets have all been cleared and the deposit has been refunded." The source supports the approval, the item count and category on the clearance sheet, and the agreed refund period. It does not state that clearance was completed or that the deposit was paid back. The draft sentence goes beyond the material.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Four libraries: closure approvals, inventory clearance sheets, lease termination agreements, asset count sheets | Check vector model identifiers for consistency first; exclude non-matching libraries before associating |
| Search mode | Semantic retrieval | Compare semantic and hybrid recall on the same natural-language question from the sample |
| Citation cap | Hypothetical starting point of 800 tokens | Compare whether long and short clearance sheet fragments stay complete, then adjust |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, then compare the cited content after filtering |
| Question optimization | Enabled | Compare the original question with the optimized question and check whether subject and store are altered |
| Result reranking | ReRank enabled | Where a model is available, run a second ordering pass on candidates and compare score-based filtering |
These values are hypothetical starting points for this exercise. They are not official defaults.
Three common errors
- Reading "the deposit is to be refunded within thirty days after handover" as a completed refund. The input is the lease termination agreement text. The comparison object is the phrase "has been refunded" in the draft sentence. The observable result is that the source only sets a refund period and does not record the refund action, so the claim is unsupported.
- Associating libraries that use different vector models. The input is the vector model identifier of each library. The comparison is whether the identifiers match. The observable result is that a library with a different identifier is excluded, and no cross-model retrieval is run to compare missed recall.
- Treating minimum relevance as a permission or entity filter. The input is the similarity threshold setting. The comparison object is whether cited content is excluded by business ownership. The observable result is that the threshold filters by relevance only and does not exclude unrelated businesses or unauthorized records.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare whether "all cleared" and "deposit refunded" have evidence with the same subject, the same period and the same field. The observable result is that when the source does not state that status, the claim is judged unsupported.
- Input a clearly labelled paper input of an empty array. Compare whether the downstream node is designed to accept an empty citation array. The observable result is that the path runs through, and no claim is made that a rule-based retrieval must return an empty array.
- Input a query containing an image. Compare whether the current vector model supports image input and whether an image index was generated at ingestion. The observable result is that when the capability is absent, the image retrieval path does not run, and the outcome is recorded as unverified.