What the documents look like
Closure accounting handles exit and asset recovery, the reverse direction from site selection and store opening. The working data comes from POS inventory, the fixed-asset system, lease contracts and financial ledgers. Typical roles include chain operations, finance, legal and the landlord contact. Everyday documents include closure approval, inventory clearance forms, lease termination agreements and asset count sheets. These materials are updated per store on an irregular schedule, and errors can lead to deposit loss, asset loss, or labour and franchise disputes.
For citation tracing, this means training and store-service answers must point back to specific closure materials. The quantity on an inventory clearance form, the signature status on a lease termination agreement, and the count scope on an asset sheet are all original-text references that must be checked line by line when answering store questions. Because updates are irregular, the same store may hold different versions at different stages, and a citation must confirm that the version used matches the current handover stage.
Constraints at the citation tracing step
The knowledge base search node can link one or more knowledge bases that share the same vector model, and it outputs a citation array whose length may be zero; an empty array still completes the output path. After retrieval results pass to the answering step, the answer text must be compared against the cited original text item by item. The citation limit controls the maximum amount of content cited per search, measured in tokens. The minimum relevance value is a 0-1 filter threshold that takes effect only in semantic retrieval or when result reranking is used. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a rerank model enabled according to settings. Image retrieval requires image input and the corresponding model capability. Semantic recall and relevance scores do not perform permission or entity isolation, and retrieval itself does not constitute authorization.
Hypothetical drill sample: a chain-brand operations specialist asks about "East China Region Store A closure handover" during internal training. The original material provided for checking is: "Inventory clearance form number QC-2024-0312 states that Store A completed its inventory count on 12 March 2024, that the difference between the physical count and the system book quantity was 3 items, and that the cause of the difference is pending investigation." "The signature status field of the lease termination agreement reads 'pending landlord signature', and the date line is blank." A draft sentence to be judged is: "The Store A inventory difference has been identified, and the lease termination agreement has taken effect." The task is to determine whether this draft sentence matches the original text.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this trial | Reason for the choice and comparison method |
|---|---|---|
| Linked knowledge bases | Select knowledge bases holding closure materials that share the same vector model | Only the same vector model can be linked together; check identifiers first and exclude non-matching bases |
| Search mode | Semantic retrieval and full-text retrieval used together | A number such as QC-2024-0312 suits full-text; a natural-language question suits semantic |
| Citation limit | Set for this trial to a token amount that can hold the original text of 2 to 3 materials | Compare against the sample material length and observe whether citations cover both original texts |
| Minimum relevance | Set for this trial to 0.5 | Compare against rerank scores and observe whether low-relevance fragments are filtered |
| Result reranking | Enabled according to settings | A rerank model is required; observe whether candidate ordering relates to question completeness |
| Citation output handover | Pass the citation array to the answering step | Check that answer sentences correspond item by item to the cited original text |
Three common errors
- Treating a retrieval hit as authorization or approval. In the sample, the lease termination agreement signature status is "pending landlord signature"; retrieving that agreement does not mean termination has taken effect, and the draft sentence claiming it "has taken effect" lacks support in the original text.
- Linking knowledge bases that use different vector models. The documents require the same vector model for joint linking; check identifiers first and exclude non-matching bases, and do not run cross-model retrieval to compare results.
- Treating minimum relevance as a permission or entity filter. That value is a filter threshold and does not exclude unrelated business or unauthorized records; retrieval does not constitute authorization.
How to verify the setup
- Input the sample original text and the draft sentence "The Store A inventory difference has been identified, and the lease termination agreement has taken effect", then compare the draft sentence against the original. The original states the difference cause is pending investigation and the agreement awaits landlord signature, so the draft sentence does not match the original and is judged unsupported.
- Input the question "What is the Store A inventory clearance form number?" and compare whether the cited content contains QC-2024-0312. If the citation array contains that number and the answer cites it, the condition is met; if the citation array is empty, record the miss and do not infer a cause.
- Input a paper empty array [] and check whether the downstream answering step handles it as designed. If the downstream design can process an empty array and give a notice that no basis was retrieved, the condition is met; this check does not claim that an actual retrieval will necessarily return an empty array.