What the documents look like
A brand concession inside a shopping mall is settled through the mall's unified cashier and a concession-fee arrangement. The brand receives sales data from the mall POS, while its own ERP records shipments and stock, and the finance system records concession fees and settlement amounts. Four material types form the comparison base: the joint-operation concession contract, which states the concession percentage and settlement cycle; the mall activity letter, which describes resource input and cost sharing for an activity period; the sales receipt, which records a single transaction; and the monthly settlement statement, which summarises period sales and the fees due.
Daily and monthly update rhythms place the same transaction at different points across the receipt, the settlement statement and the contract. Retrieval therefore has to locate material by entity, period and field separately. Values from different sources cannot be treated as one fact. A concession percentage in the contract applies to a defined settlement period, an activity letter may adjust the percentage only for stated dates, and a settlement statement aggregates a whole month.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. The node outputs references as an array, and the array length can be zero, which means the downstream chain still runs when nothing is retrieved.
Semantic retrieval uses vector distance. Full-text retrieval uses keyword matching. Hybrid retrieval combines both paths through RRF. A ReRank model performs a second ordering of candidates and returns a score between 0 and 1. The minimum relevance setting takes effect only in semantic retrieval or when result reranking is used. The reference cap is measured in tokens. Retrieval itself performs no authorization check, and a similarity score does not separate business entities or permission scopes.
Hypothetical paper exercise sample: a brand signs a joint-operation concession contract with a mall, stating that the concession rate for March 2025 is 18% of sales. An activity letter states that the concession rate is temporarily adjusted to 12% from 8 March to 10 March. March sales receipts total CNY 120,000. The monthly settlement statement records fees due of CNY 18,000. A draft sentence under review reads: "During the March activity period the concession rate is 12%, and the settlement statement shows fees due of CNY 18,000." This draft sentence mixes an activity-period rate with a whole-month settlement amount, so the contract text, the activity letter text and the settlement statement text must each be checked to confirm which period and which source value apply.
Configuration and implementation checks
| Configuration item | Setting or check for this page | Note |
|---|---|---|
| Associated knowledge bases | Contract base, activity letter base, sales receipt base, settlement statement base; confirm the same vector model | Bases on different vector models cannot be associated together |
| Search mode | Compare semantic, full-text and hybrid retrieval on the test question | No single mode is assumed to suit all material |
| Reference cap | Starting value for this paper exercise is 2000 tokens, compared against contract clauses and settlement paragraphs of different length | This value is a starting point for the exercise, not an official default |
| Minimum relevance | Mark relevant and irrelevant samples first, then adjust the threshold | Takes effect only in semantic retrieval or with result reranking |
| Question optimisation | Compare the original question with the optimised question, checking whether entity and period were altered | Used for coreference removal and question expansion |
| Result reranking | Confirm a ReRank model is configured and enabled as set | Rerank score is 0 to 1 and can be used for filtering |
Three common errors
- Treating a retrieval score as a permission filter. Minimum relevance is a filtering threshold and does not distinguish business entities or authorization scope. Retrieving a contract does not mean the contract has taken effect or that the current user is entitled to view it. Authorization status has to be checked separately.
- Comparing values from different periods or different fields directly. The 12% in the activity letter applies only from 8 March to 10 March. The 18% in the contract applies to the whole month. The CNY 18,000 on the settlement statement corresponds to whole-month sales. Each value has its own applicable period, so a difference in the numbers does not by itself show a conflict.
- Treating a pending status as approval. If the activity letter states "pending confirmation by mall operations", that status does not mean the activity has been executed or the concession rate has been adjusted. The later confirmation record has to be checked.
How to verify the setup
- Enter the contract text "the concession rate for March 2025 is 18% of sales", the activity letter text "the concession rate is temporarily adjusted to 12% from 8 March to 10 March", and the settlement statement text "fees due CNY 18,000". Compare these with the draft sentence "During the March activity period the concession rate is 12%, and the settlement statement shows fees due of CNY 18,000". If retrieval returns all three source texts and the draft sentence places the activity-period rate next to the whole-month settlement amount without marking the period difference, the condition for finding a period-mixing problem is met. The expectation is that manual review identifies the difference; no actual test is claimed.
- Enter the question "what is the concession rate for March". Compare the reference content returned by semantic retrieval with that returned by full-text retrieval. If full-text retrieval matches the contract and activity letter texts while semantic retrieval returns semantically close passages from a different period, the condition for comparing the two modes is met. A first-position match or complete recall is not guaranteed.
- Enter an empty reference array []. Check whether the downstream node continues to run. If the downstream node still carries the chain when reference length is zero, the condition for handling an empty result is met. This is a paper design check and does not claim that any rule-based retrieval would return [].
Across all three checks, relevance scores and semantic recall do not carry authorization or entity isolation. The same-vector-model requirement applies to the associated knowledge bases and does not extend to external systems. Where the source text does not state a system error or an automatic business action, no such outcome is asserted. Retrieval observations are recorded as observations, and missing evidence remains missing.