What the documents look like
Chain-store closure settlement handles exit and asset recovery, the opposite direction from head-office site selection and store opening. The material set comes from POS inventory, fixed-asset systems, lease contracts and financial ledgers. Updates are irregular and occur once per store. The routine documents are closure approval, inventory clearance forms, lease termination agreements and asset count sheets. In cross-border situations these materials may appear in different languages, so retrieval has to confirm that a cited item maps to a specific store, a specific stage and a specific version. Deposit loss, asset loss and labour or franchise disputes are the consequences of error, which is why the tracing step requires that a citation can be checked back against source text.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. It outputs cited content as an array, and the length may be zero; even when nothing is found, the output path still runs. Semantic retrieval supports cross-language understanding, and matching an English knowledge point from a Chinese question is a documented capability. The minimum relevance value is a 0-1 filtering threshold that takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold, not a final ordering rule. Hybrid retrieval merges multiple result paths through RRF, and ReRank performs a second ordering of candidate text. The citation limit is measured in tokens, not top k. Retrieval by itself does not constitute authorization: whether closure approval passed, whether a lease termination agreement was signed, and whether an asset count sheet was completed all require separate status checks.
Hypothetical drill sample: one overseas store of a chain brand enters the closure settlement stage. The source text provided for checking consists of three sentences. "The closure approval form states that the store closure application was signed off by the regional operations lead on 10 March 2025, with status pending finance review." "The inventory clearance form states that as of 15 March 2025, 120 units of category A goods in the store POS inventory had been entered as clearance quantity, while category B goods had not been entered." "The draft lease termination agreement states that both parties intend to terminate the lease on 30 April 2025; the draft has not been signed." One draft sentence to be judged reads: "The store lease was terminated on 30 April 2025, and the deposit has entered the refund process." Comparing this draft sentence with the source text: the lease termination agreement is in draft and unsigned status, and signing, reaching agreement and contract effectiveness are three distinct facts. The draft sentence turns an intended date into a completed termination and turns the deposit into an already-started refund process, while the source text records no deposit refund arrangement. The draft sentence is not supported.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Select one closure settlement knowledge base using the same vector model | The documents require the same vector model before bases can be associated together; compare identifiers against a base using a different vector model, exclude the non-matching one, and do not run cross-model retrieval |
| Search mode | Semantic retrieval | The material contains Chinese and English wording; compare hit behaviour against full-text retrieval for numbered queries, and switch numbered or model-number questions to full-text or hybrid retrieval |
| Minimum relevance | Set to 0.5 as the hypothetical starting point for this trial | The documents state the value takes effect only under semantic retrieval or result reranking; compare against 0.3 and 0.7 and observe whether returned citations include irrelevant fragments |
| Citation limit | Set to 2000 tokens as the hypothetical starting point for this trial | The documents control the citation limit in tokens; compare against 1000 and 4000 and observe whether cited content is truncated or exceeds what is needed |
| Result reranking | Enable ReRank | The documents state reranking can improve the relevance of text results; compare candidate ordering against the disabled state, and reranking requires a model and must be enabled by configuration |
| Citation output handoff | Check whether the search node output array is passed to the answering step | The documents state the output length may be zero; use a paper input of [] to check the downstream design, without claiming that a fictional rule makes retrieval always return [] |
Three common errors
- Treating the intended date in a draft lease termination agreement as a completed termination. The error is reading an intended date in an unsigned document as an effective fact. Check: input the three source sentences and the draft sentence above, compare whether the agreement status field is draft-unsigned or signed, and observe whether the draft sentence presents the intended date as a completed termination. The statement is supported only when the material clearly records that it has been signed and has taken effect.
- Treating minimum relevance as a permission or entity filter. The error is assuming that raising relevance excludes unrelated stores or unauthorized records. Check: input the same query with different relevance values, compare whether returned citations can still include material from other stores, and observe whether relevance changes the permission scope. Relevance is a filtering threshold, and retrieval does not carry out authorization decisions.
- Treating an empty citation array as a workflow failure. The error is assuming that an output length of zero means the path has broken. Check: input a paper empty array [], compare whether the downstream node still executes, and observe whether the path continues. The documents state that even when nothing is found, the output path still runs.
How to verify the setup
- Use the three source sentences and the draft sentence from the hypothetical drill sample, compare the lease status in the draft sentence against the draft-unsigned status in the source text, and observe whether the draft sentence is judged unsupported. The expected result is that the draft sentence is not supported, and the source text records no deposit refund arrangement.
- Use the same query under semantic retrieval and full-text retrieval, input a query containing a numbered question, and observe whether the numbered query is easier to hit under full-text or hybrid retrieval. The expected result is that the hit behaviour for numbered queries differs under full-text or hybrid retrieval, with no guarantee of a first-position hit or full recall.
- Use a paper input of [] to check the downstream handoff design, compare the search node output with the input of the answering step, and observe whether the empty array still lets the path continue. The expected result is that the path continues to execute; this observation is a paper check and does not claim that testing has been performed or that a fictional rule makes retrieval always return [].