What the documents look like
Pinduoduo low-price campaign operations centre on low-price resource slots, campaign inventory and refund-rate swings. The working materials for the inventory procurement and supplier records task are the campaign registration form, the product cost calculation sheet, the after-sales refund form and platform internal messages. The registration form normally states the product, the campaign price and the registration quantity. The cost calculation sheet states the purchase price, the logistics fee and the gross-margin assumption. The refund form records refund reasons and quantities. Platform internal messages carry campaign rules or penalty notices. These materials update daily and per campaign. Citation tracing requires that every procurement or supplier conclusion can be returned to the specific version of the source material, so that an older registration price does not overwrite the current campaign price.
Constraints at the citation tracing step
Citation tracing in inventory procurement and supplier records requires that each conclusion can be traced back to the corresponding source text. The knowledge base search node outputs cited content as an array, and the length may be zero. An empty array does not mean the workflow has errored or terminated. Associated knowledge bases must share the same vector model; knowledge bases with different vector models cannot be associated together. At the checking step, first compare the knowledge base identifier and the vector model identifier, exclude those that do not match, and do not run a cross-model retrieval to compare missed recalls. The minimum relevance value is a filter threshold that takes effect only in semantic retrieval or when result reranking is used, and it does not perform permission or entity isolation. The same-vector-model requirement for associated knowledge bases does not extend to external systems such as business registration sites. Full-text retrieval suits numbers, model codes and proper nouns; semantic retrieval suits natural-language questions; hybrid retrieval fuses multiple recall paths through RRF; reranking requires an available model and must be enabled in the settings. Image retrieval requires corresponding image input and model capability. The optional paths shown in the overall retrieval diagram do not mean every request executes all of them.
Hypothetical paper exercise sample: the source text reads "The 1 March 2025 version of the campaign registration form states: product A registration quantity 800 units, campaign price 9.9 yuan, registration deadline 5 March 2025. The 20 February 2025 version of the product cost calculation sheet states: product A purchase price 6.5 yuan, logistics fee 1.2 yuan, gross margin calculated at the campaign price of 9.9 yuan is 2.2 yuan." The draft sentence to be judged reads "Product A can accept 800 units of orders at the 9.9 yuan campaign price, and the purchase price of 6.5 yuan is confirmed." The draft sentence conflates the registration quantity with a confirmed purchase price. The registration form states only the registration quantity, and the cost calculation sheet gives only a calculation assumption. Neither states that the supplier has confirmed the purchase price or the orderable quantity.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge base | Select the knowledge base with the same vector model that holds the campaign registration form and the cost calculation sheet | Compare the knowledge base identifier and the vector model identifier for consistency, and exclude knowledge bases with a different vector model |
| Search mode | Hybrid retrieval | Campaign prices and product codes suit full-text retrieval, while refund reason descriptions suit semantic retrieval; compare recall differences between the two modes |
| Citation limit | Hypothetical starting point of 3000 tokens | The paper sample material is short, so this value makes it easy to observe whether the citation covers the registration form and cost calculation sheet source text; compare truncation against lower values |
| Minimum relevance | Hypothetical starting point of 0.3 | This value takes effect only in semantic retrieval or when result reranking is used; compare whether low-relevance fragments are filtered at different values |
| Question optimisation | Enable for multi-turn follow-ups | Coreference resolution in continuous dialogue can improve retrieval; compare whether the retrieval query is completed before and after enabling |
| Result reranking | Enable when candidates are numerous | Reranking requires an available model and must be enabled in the settings; compare the ordering difference between rerank scores and vector scores |
Three common errors
- Treating the registration quantity on the campaign registration form as a supplier-confirmed orderable quantity. The input is the registration form source text in the hypothetical sample above; the comparison object is the draft sentence "can accept 800 units." The registration form states only the registration quantity and does not state supplier confirmation, so the statement is unsupported.
- Treating the purchase price assumption on the cost calculation sheet as a confirmed purchase price. The input is the cost calculation sheet source text "purchase price 6.5 yuan"; the comparison object is the draft sentence "the purchase price of 6.5 yuan is confirmed." The calculation sheet gives only a calculation assumption and does not state supplier confirmation or contract effectiveness, so the confirmation is unsupported.
- Treating a retrieval result as an authorisation to execute. The input is the citation array output by the knowledge base search node; the comparison object is a procurement order placement or campaign registration submission action. The retrieval output provides cited content only. Business verification, approval and database writes require separately connected system implementation, and retrieval does not constitute authorisation.
How to verify the setup
- Check the hypothetical sample registration form source text against the draft sentence. The input is the registration form source text "registration quantity 800 units, campaign price 9.9 yuan, registration deadline 5 March 2025." Compare whether the draft sentence states the registration quantity as an orderable quantity. The observed result is that the draft sentence exceeds the source text and should be recorded as unsupported.
- Check the hypothetical sample cost calculation sheet source text against the draft sentence. The input is the cost calculation sheet source text "purchase price 6.5 yuan, logistics fee 1.2 yuan, gross margin calculated at the campaign price of 9.9 yuan is 2.2 yuan." Compare whether the draft sentence states the calculation assumption as a confirmed purchase price. The observed result is that the confirmation has no corresponding record in the source text and should be recorded as pending verification.
- Check the citation content output against the downstream handling. The input is the citation array output by the knowledge base search node; for the paper exercise it can be set to []. Compare whether the downstream answer node handles an empty array as no cited content. The observed result is that the workflow still completes without error termination. Record the actual observation rather than claiming that a fictional-rule retrieval will certainly return [].