Central chain procurement: citation tracing for after-sales questions

Review citation tracing for after-sales questions in Central chain procurement. Check input mappings, source evidence and failure handling before use.

What the documents look like

Central chain procurement work revolves around head-office purchasing and distribution. The main data sources are ERP purchasing records, WMS records, store POS records and supplier delivery data. The documents most frequently cited in order fulfilment and after-sales answers are centralised procurement contracts, replenishment plans, distribution notes and store requisition forms. These materials belong to three separate stages: purchasing, distribution and the store. For the same batch of goods, the quantity, date and responsible party recorded in the contract, the distribution note and the requisition form may differ. Before any citation is used, the document version and its valid scope must be confirmed. Data updates on a daily or weekly basis; this page does not cover batch synchronisation or real-time trigger mechanisms.

Constraints at the citation tracing step

A knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Citation content is output as an array, and the array length may be zero. When no content is found, the output path still runs through. Search results must be passed to the answering step, where each answer statement is checked against the cited source text. Minimum relevance is a filter threshold that takes effect only in semantic retrieval or when result reranking is used. It does not perform permission filtering or entity isolation. Retrieval itself does not constitute authorisation; form value collection, business verification, approval and external synchronisation each require their own downstream implementation. Hybrid retrieval merges multiple result paths through RRF, which is a fusion process and not a switch that can be turned on.

Hypothetical paper exercise sample: a chain enterprise supply chain centre receives store requisition form YA-2025-0312, dated 12 March 2025, with the requisitioning party being the East China distribution centre and the item being a certain specification of packaging material, quantity 200 boxes. The corresponding distribution note PS-2025-0315 states that 180 boxes were dispatched to that store on the same day. Centralised procurement contract HT-2025-008 stipulates that a single distribution of this specification shall not be less than 200 boxes. A draft reply reads: "On 12 March, 200 boxes were requisitioned; the distribution note shows 200 boxes dispatched; the contract stipulation has been met." This draft sentence is inconsistent with the 180 boxes in the distribution note text and also falls short of the contract stipulation. Each item must be checked before the reply is finalised.

Configuration and implementation checks

Configuration or check itemPurpose on this pageCorrespondence with the sample
Linked knowledge basesSelect the knowledge bases holding procurement contracts, distribution notes and requisition forms; confirm the vector model is the sameThe three material types are stored separately; model identifiers must be checked first
Search modeIdentifier-type queries suit full-text retrieval or hybrid retrievalDistribution note numbers and contract numbers are exact queries
Citation limitControls the amount of cited content per search, measured in tokensA hypothetical trial starting value of 2000 tokens allows comparison of whether contract clauses enter the citation in full under different limits
Minimum relevanceFilters low-relevance results; effective only in semantic retrieval or result rerankingThis sample uses identifier queries, so the threshold has limited effect
Result rerankingReranks candidates when a rerank model is available and enabledCan be enabled when the sample has many candidates, to observe clause ordering
Citation content outputOutput as an array; length may be zeroUse paper input [] to check how downstream handles an empty result

Three common errors

  1. Treating a retrieval hit as an authorisation basis. Finding the distribution note text only means the material can be recalled from the knowledge base. It does not mean distribution is complete or the contract has been performed. The check compares the quantity field in the cited source text with the draft sentence and observes whether the two agree.
  2. Treating minimum relevance as a permission or entity filter. This threshold only affects filtering of low-relevance results. It does not exclude unrelated business records or unauthorised records. The check submits the same query and compares whether returned content still includes materials from other stores or parties, observing whether out-of-scope citations appear.
  3. Forcing knowledge bases with different vector models to be linked. Cross-model linking is not permitted. The model identifier of each base should be checked first, non-conforming bases excluded and the actual observation recorded. Cross-model retrieval should not be executed to compare missed recalls.

How to verify the setup

  1. Input the requisition form text and the draft sentence from the sample. Compare the draft sentence "200 boxes dispatched" with the distribution note text "180 boxes". Observe whether the cited content matches the source text. If the citation shows 180 boxes while the draft sentence states 200 boxes, the draft sentence does not meet the checking condition.
  2. Input contract number HT-2025-008. Compare whether the cited content fully includes the stipulation "a single distribution shall not be less than 200 boxes". Observe whether this clause is truncated under the citation limit setting. If the clause is missing, adjust the citation limit and retest.
  3. Use paper input [] to check the downstream design. Observe how the answering step handles an empty citation array. An empty array does not mean the workflow has errored out, nor does it mean the retrieval rule will necessarily return an empty result. It is used only to confirm the downstream handling logic.