What the documents look like
In a hypermarket fresh-food concession counter, a supplier operates on site and shares sales revenue with the store. Procurement and supplier records therefore span four document families at once: the concession contract, the daily sales report, the shrinkage (loss) sheet, and the food incoming inspection record. These materials determine whether staff can locate the basis for three recurring jobs: reconciling weighed sales, allocating shrinkage, and checking incoming-goods certificates. At the retrieval step, the practical requirement is to tie contract clauses, batch delivery records, and shrinkage evidence to the same supplier and the same period. Where a document is absent, the retrieval result can only register a gap; it cannot stand in for an authorization decision.
A retrieval hit confirms that a document exists in the searched knowledge base. It does not confirm that a termination has taken effect, that a supplier is authorized, or that a loss-allocation clause exists. Those remain separate business facts.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Semantic retrieval, full-text retrieval, and hybrid retrieval are compared against test questions separately; no single mode is assumed to suit all four document families. The citation limit is measured in tokens, not in a guessed number of items. The minimum relevance score is a filter threshold that applies only under semantic retrieval or when result reranking is used. Question optimization performs coreference resolution and question expansion and adds an extra model call before retrieval. Image retrieval requires image input together with the corresponding model capability, and reranking requires a rerank model enabled according to settings. These are optional paths in the documented flow; their presence does not mean every request executes all of them. Retrieval results do not constitute authorization, and they do not perform entity isolation.
Hypothetical drill sample: a concession contract text states that "Party A is the fresh-food department of a hypermarket; Party B is a vegetable supplier; the contract is valid from 1 January 2025 to 31 December 2025; Party B operates on site in Party A's fresh-food area and shares revenue at 15 percent of sales." A draft handover sentence reads: "This supplier's contract has been terminated, and inventory shrinkage is borne by the store." The request to be judged is: "Please retrieve the basis for this supplier's shrinkage allocation." The source text contains the subject, the object, and the validity range. It does not state a termination fact and does not state a shrinkage-bearing clause.
Configuration and implementation checks
| Configuration item | Setting for this page | How to check |
|---|---|---|
| Linked knowledge bases | Link knowledge bases containing the concession contract, daily sales report, shrinkage sheet, and food incoming inspection record | Check the vector model identifier of each base and exclude any that do not match |
| Search mode | Run semantic retrieval and full-text retrieval separately against the sample question | Compare the returned content from each path against the subject of the question |
| Citation limit | For this hypothetical trial, start at 2000 tokens | This value is used to observe whether long and short materials are truncated, and is compared against shorter materials |
| Minimum relevance | For this hypothetical trial, start at 0.3 | First label related and unrelated samples, then compare results on both sides of the threshold |
| Question optimization | Enable for sample questions containing coreference | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable according to settings when a rerank model is available | Compare candidate order before and after reranking, and record what is observed |
Three common errors
- Treating a retrieval hit as authorization. In the sample, the draft sentence states the contract has been terminated, while the source text states only a validity range and no termination fact. Retrieving the contract text shows that the material exists; it does not show that a termination has taken effect.
- Treating the minimum relevance score as an entity filter. That value is a filter threshold; it does not exclude unrelated suppliers or unauthorized records. The sample request still requires a separate check of the supplier identifier and access permission.
- Describing RRF as a switch. RRF is a fusion process in hybrid retrieval, not an item that can be turned on. In the sample, the semantic and full-text result paths are handled through fusion, and the first hit is not guaranteed.
How to verify the setup
- Input the sample contract text and the draft sentence, then compare the draft sentence's claim of "has been terminated" against the validity range field in the source text. The source text does not state a termination, so the claim is recorded as unsupported, and the evidence gap is logged.
- Input the sample request "Please retrieve the basis for this supplier's shrinkage allocation," then compare whether the returned content contains a shrinkage-bearing clause. The source text does not state such a clause; when the return is empty or contains only contract validity content, record that no basis was found.
- Input an explicit paper empty result
[], then check whether the downstream design accepts an empty citation. The node output citation length can be 0, and the empty-result path still completes; this observation does not support inferring that a retrieval rule must return empty.