Wholesale stall trading: citation tracing for product questions

Review citation tracing for product questions in Wholesale stall trading. Check document scope, failure paths and source evidence through controlled tests.

What the documents look like

Wholesale market stall trading is fragmented. Prices are negotiated on the spot, and goods move by consignment. The paperwork that matters here is the purchase receipt, the handwritten sales slip, the stall lease contract, and the payment receipt. Each of these records a different fact, and the same product can change source document from one order to the next because updates happen daily and per order. A product question therefore has to point to the specific slip that supports the answer.

The documents are brief and often handwritten. The value of citation tracing in this setting is that it binds an answer to a checkable original text instead of to a general impression of what the stall usually sells. For product information and selection questions, the answer is only as good as the document it cites.

Constraints at the citation tracing step

The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its citation output is an array, and the length can be zero, meaning the output path still runs even when nothing is retrieved. The retrieved citations then pass to the answering step, where each answer sentence is checked against the cited original text.

The minimum relevance setting takes effect only under semantic retrieval or result reranking. It is a filter threshold and does not perform permission or entity isolation. The citation limit is counted in tokens, not in top k. Hybrid retrieval merges results through RRF, and reranking requires a rerank model and must be enabled according to settings.

A relevance score is not a category, entity, or permission filter. Retrieval returns citations; it does not decide whether a buyer may collect goods or whether payment has been received. Those are separate business facts that need their own downstream implementation.

Hypothetical exercise sample: a stall keeps a handwritten sales slip whose text reads "12 March 2024, sold to buyer A, rebar HRB400 diameter 20 mm, quantity 3 tonnes, unit price at market, payment not received." A separate stall lease contract states a term from 1 January 2024 to 31 December 2024. A draft sentence reads "This batch of rebar has been paid for and can be supplied under the contract until the end of 2024." When checked, the sales slip states payment not received, while the draft sentence states payment received. These are mutually exclusive values in the same field, so this is a contradiction. The contract term does not represent a supply period, so the draft sentence treating the lease term as a supply period is unsupported.

Configuration and implementation checks

Configuration itemHypothetical starting point for this exerciseReason and comparison method
Associated knowledge basesOnly two bases sharing the same vector modelThe document requires the same vector model before bases can be associated together; check the identifier first and exclude those that do not match
Search modeHybrid retrievalNumber and model questions need full-text recall, while natural language questions need semantic recall
Minimum relevance0.3Hypothetical starting point; compare with 0.5 and observe changes in the number of retrieved items; not an official default
Citation limit2000 tokensHypothetical starting point; compare with 1000 and observe the number of citations; not an official default
Result rerankingEnabledSecondary ranking when candidates are numerous; requires a rerank model
Question optimizationEnabled for multi-turn follow-upsAdds one model call; can be enabled for short-question scenarios

Three common errors

  1. Treating retrieval as authorization. Retrieval only returns citation content and does not determine whether buyer A is entitled to collect goods or whether payment has been received. Check: input "can buyer A collect the goods," compare whether the answer cites the sales slip line "payment not received," and observe whether the answer makes a conclusion beyond its authority.
  2. Treating the lease term as a supply period. The contract term and the supply period are different fields. Check: input the draft sentence, compare against the contract text that states only a lease term, and observe whether the lease term is presented as a supply period.
  3. Treating an empty citation as an error. The citation array length can be zero, and the output path still runs. Check: input a paper empty array, compare against the downstream design for accepting empty results, and observe whether it is misjudged as a retrieval failure.

How to verify the setup

  1. Input the sales slip original text and the draft sentence. Compare "payment not received" with "has been paid for." The condition is met when the check identifies these as mutually exclusive values in the same field and marks the draft as contradictory.
  2. Input the contract original text and the draft sentence. Compare the lease term with the supply period. The condition is met when the check marks the supply-period claim as unsupported because the contract states only a lease term.
  3. Input a paper empty array. Compare against the downstream acceptance design. The condition is met when the empty result passes through normally and is not reported as an error.

These checks describe expected observations and do not claim that any test has been run. Missing evidence remains missing. A pending review does not demonstrate a lack of operating authority. Signing, reaching agreement, and contract effectiveness are three distinct facts. A quotation validity period does not indicate a supply period. Where the source documents do not state a term, the evidence gap is recorded rather than filled with an invented authorization or effectiveness clause. The materials needed for verification must sit within an accessible target retrieval scope or be supplied as explicit manual input; when they are absent, the gap is recorded.