What the documents look like
Supermarket supplier onboarding and settlement is commonly handled as a purchase-and-sale account or a consignment account, and a resident on-site operation is not necessarily part of the arrangement. The working documents are purchase contracts, delivery acceptance sheets, rebate agreements and reconciliation statements. These come from ERP purchasing, the supplier portal, POS sales and the finance system, and their update rhythm is weekly or monthly.
For product information and selection questions, the questions to answer usually point at one supplier, one settlement period, one acceptance sheet or one rebate clause. A citation source has to lead back to the matching material. When citation tracing is weak, the visible consequences are account disputes, a supplier stopping delivery, and friction with fair-trading requirements.
Constraints at the citation tracing step
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; an empty array still lets the output path run through. The citation cap is controlled in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Neither the citation cap nor the minimum relevance value performs permission isolation or entity isolation. Whether a retrieved result is usable depends on whether the material has finished training and whether its version matches the current settlement period.
Hypothetical paper exercise: a supermarket and a supplier sign a March 2025 purchase contract. A contract annex states a settlement deduction rate of 3% for a named product. An April 2025 delivery acceptance sheet records 200 units arrived for that product and 198 units accepted as qualified. A draft sentence reads: "The April settlement deduction rate for this product is executed at 5%." The pending request reads: "Please confirm the April deduction rate against the contract and the acceptance sheet."
The contract and the acceptance sheet are the materials to compare. The draft sentence and the pending request are the items awaiting judgment. No approval, settlement execution or database write has occurred at this stage.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | How to check it |
|---|---|---|
| Associated knowledge bases | Select the contract base and the acceptance sheet base that share one vector model | Compare the selected base identifiers and confirm the vector model is the same |
| Search mode | Hybrid search is the hypothetical starting value | Compare against semantic search on numbered and model-type queries |
| Citation cap | 3000 tokens is the hypothetical starting value | Compare against a lower value and observe how the number of cited items changes |
| Minimum relevance | 0.3 is the hypothetical starting value | Compare against 0.5 and observe whether low-relevance fragments are filtered |
| Question optimization | Off is the hypothetical starting value | Compare against on and observe whether citations change in follow-up turns |
| Result reranking | Not enabled is the hypothetical starting value | Compare against enabled and observe whether candidate ordering changes |
The starting values above are chosen to give the exercise a fixed reference point, not because any of them is an official default. Each is compared against a second value so the direction of change can be observed. RRF is a fusion process applied to mixed retrieval results; it is not a switch to turn on. A thinking budget is a reasoning budget and is not the citation token cap.
Three common errors
- Carrying the March contract deduction rate straight into April settlement. The contract period is March 2025, the April acceptance sheet does not state a deduction rate, and the 5% in the draft sentence lacks evidence from the same period.
- Treating a retrieval hit as an authorization to release. The search node outputs a citation array only. Settlement approval and database writes require downstream implementation.
- Treating an empty citation array as a workflow failure. The output length may be zero, and the downstream step should accept an empty array and prompt for additional material.
How to verify the setup
- Input the contract text and the acceptance sheet text from the hypothetical exercise, then compare the deduction rate in the draft sentence against the deduction rate in the contract annex. The contract states 3% and the draft sentence states 5%. These are mutually exclusive values on the same field, so the comparison result is a contradiction.
- Input the pending request, then compare the subject and period of the request against the subject and period of the cited source text. The request is for April, the contract is for March, and the acceptance sheet does not state a deduction rate. The result is unsupported, and the evidence gap is recorded as missing.
- Input a clearly labelled paper empty array, then compare whether the downstream node continues and emits a prompt. If the downstream node accepts the empty array and prompts for additional material, the check passes. If the downstream node errors out and stops, the check fails.
Before any of these checks, confirm that the associated knowledge bases share the same vector model. Bases with different vector models cannot be associated together. Record the identifiers, exclude the ones that do not match, and note what was observed. Do not run a cross-model retrieval to compare missed recalls. An image retrieval path requires an actual image input and a model with the corresponding capability. Reranking requires a rerank model and must be enabled in the settings. The optional paths shown in the overall retrieval diagram do not all execute on every request.