What the documents look like
Brand-owned stores are staffed and accounted for directly by the brand, so their evidence chain for contract and transaction material comparison comes from the POS system, member CRM, ERP inventory, and store inspections rather than from a franchise agreement. The everyday documents are daily sales reports, shift schedules, display photos, and inventory transfer orders, refreshed daily or weekly. These materials are numerous and scattered: one transaction can leave different versions of itself in POS settlement records, transfer orders, and inspection notes. Citation tracing requires every conclusion to point back to a specific material, a specific field, and a specific point in time. Without that, a gap between a daily sales report and an inventory transfer order cannot be located to a responsible step.
Two features of this category shape the task. First, the comparison object is a draft sentence or handover request judged against source material, not a free-form summary. Second, the material set includes documents whose status is still in progress, so a field value such as pending review is a fact about the document, not evidence that an operation was performed or approved.
Constraints at the citation tracing step
The knowledge base search node outputs a citation array, and its length can be zero. An empty array means the output path runs through; it does not mean the workflow terminated with an error. Before retrieval, check which knowledge bases are selected and which material version they hold. Only knowledge bases sharing the same vector model can be associated together. Semantic retrieval computes similarity through vector distance, full-text retrieval depends on keyword hits, hybrid retrieval fuses multiple result paths through RRF, and a rerank model produces a 0-1 score used for filtering. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. A retrieval action by itself is not authorization, and a form collecting values is not a completed approval.
Hypothetical drill sample. Material text one, daily sales report excerpt: "On 10 March 2025, the POS end-of-day summary for brand-owned store A shows daily sales of RMB 12,400 across 86 transactions, sourced from the POS system end-of-day settlement." Material text two, inventory transfer order excerpt: "Transfer order number DB-20250310-07, outbound store A, inbound store B, item SKU-4471, quantity 30 units, prepared on 10 March 2025, status pending review." Draft sentence to be judged: "Store A completed the outbound transfer of 30 units on 10 March 2025, and the sales and inventory data are consistent." This draft treats a pending-review status as a completed outbound transfer and supplies no comparison field showing sales data and inventory data are consistent.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Select the daily sales report base and the inventory transfer order base | Both bases must share the same vector model; check each base identifier, exclude any with a different vector model, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Transfer order numbers are exact identifier queries while daily sales reports are natural-language descriptions, so semantic recall and full-text recall are both used |
| Minimum relevance | Hypothetical starting value 0.35 | A starting point for this drill, not an official default; compare against recall counts with no threshold to observe whether unrelated fragments decrease |
| Citation cap | Hypothetical starting value 3000 tokens | A starting point for this drill, not an official default; compare against 1000 tokens to observe whether key fields are truncated |
| Result reranking | Enabled | With many candidates, text is reordered a second time and the rerank score is used for filtering; compare against ordering with reranking disabled |
| Question optimization | Not enabled in this drill | Transfer order numbers and dates are explicit queries, so coreference resolution should not alter identifier meaning; compare against the retrieval terms when enabled |
Three common errors
- Treating a pending-review transfer order as a completed outbound transfer. The draft sentence states "completed the outbound transfer of 30 units," while material text two carries the status value "pending review." The input is the draft sentence and the transfer order text; the comparison is the status field value; an outbound conclusion is supported only when the status is explicitly completed.
- Treating a retrieval hit as an authorization basis. Retrieving the transfer order text shows only that the document exists; it does not show that the transfer was approved. The input is the citation array returned by retrieval; the comparison is whether it contains an approval field; when the approval record is absent, the conclusion stays pending verification.
- Forcing association of knowledge bases with different vector models. The input is the vector model identifier of each knowledge base; the comparison is whether the identifiers match; when they do not, exclude that base and record the observed gap rather than inferring missed recall or redundant results.
How to verify the setup
- Input material text one, material text two, and the draft sentence from the hypothetical sample, then compare the draft phrase "completed the outbound transfer" against the transfer order status field "pending review." Condition satisfied: the transfer order status field is fully recalled in the citation content, and the verification conclusion is marked as unsupported.
- Input a paper empty array [] as the retrieval output, then compare whether the downstream answer step accepts it normally. Condition satisfied: the downstream node continues after receiving the empty array without an error termination and without fabricating citation content.
- Input the vector model identifiers of the daily sales report base and the inventory transfer order base, then compare whether they are the same. Condition satisfied: matching identifiers allow joint association; non-matching identifiers lead to excluding that base and recording the gap, without running cross-model retrieval to compare missed recall.
These checks observe and describe failure branches; they do not guarantee a first-hit result, full recall, or that an unrelated query returns nothing. The empty-array check uses an explicit paper input, and does not claim that a fictional rule makes retrieval return an empty array. Cross-model bases cannot be associated, so the check confirms identifiers and excludes non-matching ones rather than running cross-model retrieval. Optional paths in the overall retrieval flow, such as image retrieval or reranking, run only when the corresponding input and model capability exist and the setting is enabled.