What the documents look like
Chain-store closure settlement handles exit and asset recovery, running in the opposite direction from site selection and store opening. The main data sources are POS inventory, the fixed-asset system, lease contracts and financial ledgers. The documents handled day to day are closure approvals, inventory clearance forms, lease termination agreements and asset count sheets. These materials sit across different systems and offline signing steps. When order fulfilment and after-sales answers cite them, unclear source labelling widens the consequences of deposit loss, asset loss or franchise disputes, because the traceable basis is thin. Updates are irregular and per store. This page does not cover batch synchronisation or real-time triggers.
The knowledge-base search node can associate one or more knowledge bases that share the same vector model. Its citation output is an array, and the length may be zero; when the array is empty, the output path still runs through. After retrieval results reach the answering step, each answer sentence is checked against the cited source text. A reasoning process or a fluent reply does not substitute for source evidence. Semantic recall and relevance scores serve ordering and filtering only; they do not carry permission or entity isolation, and retrieval itself is not authorisation. Material timeliness is tested with the current version and a historical version of a document, and whether a citation is adopted still depends on review of the effective version.
Constraints at the citation tracing step
The search node takes an input of associated knowledge bases and an input of search parameters. The citation output is an array whose length may be zero. The node is a core module, can be added repeatedly, has external input, has static configuration, and triggers execution. These are the documented behaviours available for citation tracing. Anything beyond them, such as an approval step, a database write or an external synchronisation, needs its own downstream implementation and is not implied here.
Hypothetical drill sample, clearly labelled as invented for this page: a chain-brand operations staff member handles the lease handover of a signed store. The supplied source text reads: "Lease termination agreement (draft) states: Party A and Party B reach an intention to terminate the premises at No. 12, a certain road; the signing date is blank; neither party's seal field is filled in." The draft sentence to be judged reads: "The store lease has been terminated and the deposit has been refunded." The question is whether that draft sentence has source support.
Configuration and implementation checks
| Check item | Hypothetical starting value for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Only bases holding closure-settlement material with the same vector model | Compare vector-model identifiers across existing bases and exclude different-model bases |
| Citation cap | Set to 2000 tokens | Compare against the total length of the sample material to confirm coverage of the termination agreement and count sheet |
| Minimum relevance | Set to 0.5 | Compare against semantic retrieval scores and observe whether low-scoring fragments are filtered |
| Search mode | Hybrid search | Compare against pure semantic retrieval results and check hits on numbered fields |
| Question optimisation | Off | Compare against the time cost when on; the sample question has a complete subject and needs no completion |
| Citation display | Show source text | Compare against the answer text and check each citation against the source line by line |
These values are hypothetical starting points for this drill, not official default values. The citation cap uses the token unit; thinking budget is a reasoning budget and must not be moved into the citation token cap. RRF is a fusion process for mixed results, not a switch that can be turned on. The documented search parameters also include search mode, search filtering and question optimisation, and the minimum relevance value takes effect only in semantic retrieval or when result reranking is used.
Three common errors
- Treating a retrieval hit as completed authorisation. Observable error: writing "the lease has been terminated" directly from "a draft termination agreement was retrieved". Check: input the sample source text, compare the draft sentence against the signing-status field of the source, and when the signing date is blank and the seal fields are unfilled, the draft sentence is unsupported.
- Treating an empty citation output as a workflow failure. Observable error: judging the workflow as errored when the citation array is zero. Check: input a paper-level empty array [], compare the downstream node design, and confirm the output path still runs through on an empty array.
- Associating knowledge bases with different vector models. Observable error: selecting bases built on different vector models together. Check: verify each base's vector-model identifier, exclude those that do not match, and record the actual observation. Do not run cross-model retrieval to compare missed recall.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence. Compare the draft sentence "the lease has been terminated and the deposit has been refunded" against the source text "the signing date is blank; the seal fields are unfilled". The expected result is that the draft sentence is unsupported and the evidence gap is recorded.
- Input a paper-level empty array []. Compare how the downstream node is designed to accept an empty citation. The expected result is that the output path runs through and is not judged as an error. Do not claim that a fictional rule makes retrieval always return [].
- Input a retrieval request containing the termination agreement and the asset count sheet. Compare the returned citations against the source text item by item. The expected result is that citations trace back to specific material; any citation with no matching source text is recorded as pending verification.