Supermarket supplier accounts: retrieval for operating reviews

Review retrieval for operating reviews in Supermarket supplier accounts. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Supermarket supplier settlement for store entry runs mainly on purchase-and-sale or consignment payables, and the supplier does not necessarily operate a staffed counter on site. This differs from a large-store joint counter arrangement. The main data comes from ERP purchasing, the supplier portal, POS sales, and the finance system. Everyday documents include purchase contracts, delivery acceptance notes, rebate agreements, and reconciliation statements. Operating reporting and exception review must check accepted quantities, rebate deduction rates, and return discrepancies. Materials update weekly or monthly. The retrieval and recall step works on the text content of these documents, and its results support manual checking only. Retrieval results do not constitute account confirmation or authorization to execute.

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. Search runs against the whole knowledge base; collections serve only to classify and manage data and do not affect search results. A single retrieval combines semantic search, full-text search, hybrid search, question optimization, and reranking according to the input content and search parameters, and finally fuses the results into reference content. Reference content is output as an array and its length can be 0, which means the output path still runs even when nothing is found. The minimum relevance value takes effect only under semantic search or when result reranking is used, and that value is a filter threshold. It does not carry out permission or entity isolation. Retrieval does not perform authorization, approval, or database writes.

Hypothetical drill sample: a supplier signs a purchase contract with a supermarket, and the contract states a settlement cycle of monthly settlement at 30 days. A delivery acceptance note states receipt of 100 cases on 5 March 2025. A rebate agreement states a quarterly rebate of 2%. A reconciliation statement draft reads "March 2025 accepted quantity 100 cases, rebate calculated at 2%, payment made on 4 April 2025." The draft sentence to be judged is "Payment was made under the contract on 4 April 2025." The payment fact in that sentence is not stated in the purchase contract, the delivery acceptance note, or the rebate agreement, so it remains unverified. The retrieval scope is limited to the knowledge base holding these four documents.

Configuration and implementation checks

Configuration itemSetting for this pageNote
Associated knowledge basesThe knowledge base holding the purchase contract, delivery acceptance note, rebate agreement, and reconciliation statementEach knowledge base must use the same vector model; knowledge bases on different models are not associated together
Search modeCompare semantic search, full-text search, and hybrid search separately for each test questionNo single mode is presumed suitable for all materials
Reference limitHypothetical trial starting value of 2000 tokensChosen because each of the four documents has short passages; adjust after comparing content completeness against a smaller value
Minimum relevanceHypothetical trial starting value of 0.3Takes effect only under semantic search or result reranking; label relevant and irrelevant samples first, then adjust
Question optimizationEnabled for short questions such as "was that payment made"Used for coreference resolution and question expansion; compare the subject in the original question and the optimized question
Result rerankingEnabled or not according to actual model capabilityThe rerank score is used for filtering and does not change the authorization boundary of retrieval

Three common errors

  1. Treating a retrieval result as an authorization basis. Retrieving the draft reconciliation statement line "payment made on 4 April 2025" does not support a finding that payment was completed. Payment is a downstream finance system action, and the retrieval step only returns reference content. When checking, confirm the status field of that sentence in the source document. When the payment fact is not stated, record it as unverified.
  2. Treating minimum relevance as an entity or permission filter. Minimum relevance is a 0-1 filter threshold and does not exclude unrelated business or unauthorized records. If the test question is "March 2025 accepted quantity," retrieval may return a rebate agreement fragment. Compare by hand whether the returned content matches the fields of the question, and do not assume the threshold has confined results to accepted quantity.
  3. Associating knowledge bases that use different vector models. The knowledge base search node requires associated knowledge bases to use the same vector model. If a purchase contract knowledge base and a reconciliation statement knowledge base use different vector models and are associated at the same time, first check the vector model identifier of each knowledge base, exclude those that do not conform, and record the actual observation. Do not run a cross-model retrieval to compare missed recall.

How to verify the setup

  1. Input the purchase contract text "settlement cycle of monthly settlement at 30 days" and the delivery acceptance note text "receipt of 100 cases on 5 March 2025" from the hypothetical sample, and set the search question to "March 2025 accepted quantity." Compare whether the returned references contain the 100-case record from the delivery acceptance note and whether the 2% clause from the rebate agreement is mixed in. The condition is met when the returned content contains the accepted quantity field and that field matches the delivery acceptance note text. If the returned content contains only the rebate rate, record it as a miss on accepted quantity and check the search mode and reference limit.
  2. Input the draft sentence to be judged, "Payment was made under the contract on 4 April 2025," and set the search question to "payment on 4 April 2025." Compare whether the returned references contain a source document field stating the payment fact. The condition is met when the returned references explicitly state the payment date and payment status. If the returned content contains only the contract settlement cycle or the drafted sentence in the reconciliation statement, record it as unsupported, keep the evidence gap, and do not find that payment was completed.
  3. Input an empty array [] as the reference content and check whether the downstream design handles a length of 0. Compare whether the downstream node continues to execute when references are empty and whether it mislabels the empty result as "no related records." The condition is met when the downstream path runs through with empty references and does not produce a negative conclusion. If the downstream design depends on non-empty references, record it as requiring separate implementation and do not claim that the retrieval rule returns an empty array.