Supermarket supplier accounts: citation tracing for copy drafting

Review citation tracing for copy drafting in Supermarket supplier accounts. Check sources, input boundaries, error handling and repeatable test evidence.

What the documents look like

Supermarket supplier onboarding and settlement usually runs through purchase-sale or consignment payables, and a supplier may have no resident staff on site. Four document families carry most of the evidence for product and marketing copy: the purchase contract, the delivery and acceptance sheet, the rebate agreement, and the reconciliation statement. These update weekly or monthly, so a copy draft can easily cite a figure that belongs to an earlier settlement period.

For this category, product copy that states price, rebate percentage, settlement timing, or supply commitments has to correspond line by line to those materials. Settlement disputes, supplier suspension of delivery, and fair-trading requirements make source tracing a necessary drafting step rather than an optional polish. The drafting task here is narrow: given specific source sentences from the materials, decide whether a proposed copy sentence is supported, unsupported, or still awaiting verification.

Constraints at the citation tracing step

The knowledge base search node associates one or more knowledge bases that share the same vector model, and outputs cited content as an array whose length may be zero. A zero-length array means the output path still completes; it does not mean the workflow has errored out or terminated. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result re-ranking is used, and it does not perform permission or entity isolation. The citation limit is measured in tokens and is unrelated to any thinking budget. Hybrid retrieval merges multiple retrieval paths through RRF, and re-ranking filters by a 0–1 score. Source-currency checking has to separate the current version of a document from a historical version.

Hypothetical drill sample. Source sentence one: "Party A, XX Supermarket, and Party B, YY Foods, signed a purchase contract on 1 March 2025, covering supply from April to June 2025, with rebates settled monthly at 2% of the accepted amount." Source sentence two: "The delivery and acceptance sheet dated 12 April 2025 records that YY Foods delivered 200 boxes of biscuits, of which 180 boxes passed acceptance." The proposed copy sentence under review: "YY Foods commits to a monthly rebate of 3% from April to June, settled on acceptance."

The proposed sentence conflicts with source sentence one on the rebate percentage and on the settlement cycle, and its acceptance-settlement claim has no support in source sentence two, which records a delivered quantity of 200 boxes against 180 boxes that passed acceptance. The materials do not state whether any rebate was actually settled, so that remains an evidence gap.

Configuration and implementation checks

Configuration itemHypothetical starting point for this drillReason and comparison method
Associated knowledge basesSelect the purchase contract base and the acceptance sheet base that share one vector modelCompare base identifiers; bases with different vector models cannot be associated together
Search modeHybrid retrievalCompare against semantic retrieval and full-text retrieval results, checking hits on numbers and proper nouns
Minimum relevanceHypothetical starting point 0.5Compare with 0.3 and 0.7, observing changes in cited content
Citation limitHypothetical starting point 2000 tokensCompare with 1000 and 4000, checking citation count and truncation
Result re-rankingEnabledCompare with disabled, checking the 0–1 score ordering
Citation handoffRetrieval output passed to the answering stepCheck each answer sentence against the cited source text

These starting values are drill settings chosen to give the comparison a fixed reference point. They are not documented defaults, and no official limit, unit, or menu is implied by them.

Three common errors

  1. Reading source sentence one's "rebates settled monthly at 2% of the accepted amount" as "monthly rebate of 3%, settled on acceptance." Input: the two source sentences and the proposed copy sentence. Compare the rebate percentage, the settlement cycle, and the acceptance condition. Observable result: the proposed sentence's rebate percentage and settlement cycle do not match source sentence one, and its acceptance-settlement wording has no basis in source sentence two.
  2. Expanding source sentence two's "180 boxes passed acceptance" into "all 200 delivered boxes passed acceptance." Input: the acceptance sheet sentence. Compare the delivered quantity with the accepted quantity. Observable result: the two figures differ, and the proposed expansion is unsupported.
  3. Treating the minimum relevance value as a permission or entity filter, on the assumption that a low score excludes unrelated suppliers. Input: the search parameters and the returned citations. Compare the score with the record's business ownership. Observable result: the score does not change which business a record belongs to, and it cannot be used to exclude unauthorized records.

How to verify the setup

  1. Input source sentence one, source sentence two, and the proposed sentence "YY Foods commits to a monthly rebate of 3% from April to June, settled on acceptance." Compare the proposed rebate percentage and settlement cycle against source sentence one, and the acceptance-settlement claim against source sentence two. The expected observation: the rebate percentage and settlement cycle do not match source sentence one, and the acceptance-settlement claim lacks support in source sentence two, so the proposed sentence is judged unsupported.
  2. Input the identifiers of two knowledge bases that share one vector model, together with a search request. Compare whether the identifiers indicate the same vector model. The expected observation: association is permitted when the identifiers match; when they do not match, exclude that base, record the actual observation, and do not run a cross-model retrieval.
  3. Input a paper-level empty citation array [] together with the downstream answering step design. Compare whether the downstream step completes its path on the empty array. The expected observation: the downstream step does not error out or terminate and produces no cited content, which satisfies the empty-result handoff condition; if the downstream step errors, record that gap.

Across all three checks, a pending review is not an approval, an offered option is not an execution, and an unsigned termination is not a completed termination. Where the materials are silent, the status remains awaiting verification rather than being treated as a conflict or a denial.