What the documents look like
Supermarket supplier entry settlement revolves mainly around purchase-and-sale or consignment payables. A supplier does not necessarily operate a staffed counter on site, and the settlement basis is spread across purchase contracts, delivery acceptance notes, rebate agreements, and statements of account. Materials update weekly or monthly. Retrieval needs to place materials for the same supplier, the same period, and the same field together for cross-checking. A purchase contract states rebate deduction points, a delivery acceptance note records the quantity actually received, and a statement of account aggregates the amount payable. The field definitions across these three material types do not match, so retrieval results are candidate content only and do not constitute account confirmation or authorization.
For this task, three industry facts shape the retrieval step. First, the settlement basis is fragmented across contract, acceptance, rebate, and statement documents, so a single query rarely returns a complete evidence set. Second, the same supplier and the same settlement period must be aligned before any field comparison is meaningful, which makes entity and period matching a prerequisite rather than a retrieval feature. Third, the materials update weekly or monthly, so a retrieved candidate may reflect an earlier revision than the one currently under review.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. The smallest search unit is the knowledge base; a collection is used only for classification management and does not affect search results. Cited content is output as an array and its length can be 0, which means the output path still runs even when no content is found.
The minimum relevance value applies only under semantic retrieval or when result reranking is used. It is a filter threshold and does not perform permission filtering or entity isolation. Hybrid retrieval merges multiple result paths through RRF, and reranking requires a model and must be enabled according to settings. Image retrieval requires corresponding image input and model capability. These conditions shape the candidate range for this task; retrieval itself does not complete account reconciliation.
Assumed drill sample: a supplier signs a purchase contract with a supermarket. The contract text states "the rebate is calculated at 2% of the monthly tax-inclusive purchase amount, and the settlement period is March 2024." The delivery acceptance note for the same period records "1,200 items arrived in March, of which 1,180 items passed acceptance." A draft statement of account states "the March rebate payable is calculated on 1,200 items." The draft sentence to be judged is "the March rebate should be calculated on the purchase amount corresponding to the 1,180 accepted items." The contract and the acceptance note do not state mutually exclusive values for the same rebate calculation base field. This draft sentence is unsupported and requires the original rebate agreement text before it can be verified.
Configuration and implementation checks
| Configuration item | Assumed trial starting point | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | Purchase contract base, acceptance note base, statement base | Confirm the three bases use the same vector model identifier, then associate them together |
| Search mode | Hybrid retrieval | Contract numbers and rebate ratios are exact fields; semantic recall is retained at the same time |
| Citation limit | Assumed starting point of 3000 tokens for this trial | Statement paragraphs are long; compare with 1500 tokens for content completeness |
| Minimum relevance | Assumed starting point of 0.35 for this trial | Label relevant and irrelevant samples first, then compare filtering effect with 0.5 |
| Question optimization | Enabled for multi-turn follow-up | Compare the original question with the optimized question and check whether the supplier entity was altered |
| Result reranking | Enabled when candidates are numerous | Rerank score is 0-1; compare with semantic score for ranking stability |
Three common errors
- Treating a retrieval hit as account confirmation. Retrieval returns candidate citations. Rebate amounts, accepted quantities, and settlement status require manual comparison against original materials. Retrieval does not execute approval or database writes.
- Using minimum relevance to exclude unrelated suppliers. This value is a similarity filter threshold and does not filter by supplier or permission. Unrelated records may still be recalled and require manual entity checking.
- Treating a statement under review as confirmed. The draft statement of account is in a pending-review state. It does not indicate that the rebate has been assessed, nor that the supplier has confirmed it. The evidence gap must be preserved.
How to verify the setup
- Enter the contract text, the acceptance note text, and the draft statement above. Compare the supplier name, settlement period, and rebate field across the three materials. Observe whether the contract and acceptance note for the same supplier and the same period are recalled. If only materials from other periods are recalled, record this as a miss for that period.
- Enter the draft sentence "the March rebate should be calculated on the purchase amount corresponding to the 1,180 accepted items." Compare the wording of the contract and the acceptance note regarding the rebate base. Observe whether mutually exclusive values appear for the same field. When no mutually exclusive values appear, record this as unsupported rather than as a contradiction.
- Enter the explicit paper-level empty result
[]to check the downstream design. Observe whether the output path still runs and confirm how the downstream handles an empty citation array. This check does not mean that retrieval of a fictional rule will necessarily return an empty array.