What the documents look like
Pinduoduo low-price campaign store operations revolve around campaign resource slots, campaign inventory, and refund-rate swings. The working materials are campaign registration forms, product cost calculation sheets, after-sales refund records, and platform in-site messages. The most time-consuming steps are estimating campaign inventory, tracking abnormal refunds, and checking campaign-price gross margin. Copy drafting has to check the price and inventory basis in the cost sheet, the registration conditions and campaign scope in the registration form, the refund-reason distribution in the refund records, and the campaign rules and deadlines in the in-site messages. These materials together decide whether a promotional sentence has a corresponding basis.
For retrieval, the practical consequence is that one draft sentence can draw on several documents with different formats and different validity ranges. A campaign-price figure sits in a cost sheet, a deadline sits in a registration form, and a shipping or stock promise may appear in no document at all. Retrieval returns candidate text; it does not confirm that the candidate text covers the claim being drafted.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases with different vector models cannot be linked together. At implementation, check the vector model identifier first, exclude those that do not match, and record what was actually observed. Do not run a cross-model search to compare missed recalls.
Search mode is compared per test question across semantic retrieval, full-text retrieval, and hybrid retrieval. No single mode is assumed to suit all materials. The citation limit is measured in tokens, so test with actual materials of differing length and check whether content stays complete before adjusting the limit. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; label relevant and irrelevant samples first, then adjust the threshold. Question optimization performs coreference resolution and question expansion, so compare the original question with the optimized question and check whether the subject was altered. Hybrid retrieval merges multiple result paths through RRF. Reranking requires a model and must be enabled by configuration. Image retrieval requires image input and the corresponding model capability. These paths do not all execute on every request. Cited content is output as an array and its length can be 0, meaning the downstream path still runs when nothing is retrieved.
A hypothetical paper exercise sample: the source material reads "The registration deadline for this campaign is 12 March at 18:00, the resource slot is flash sale, and registered products must hold the campaign price from 0:00 on 13 March to 24:00 on 15 March." It also reads "The product cost calculation sheet states: campaign price 29.9 yuan, unit cost 24.5 yuan, projected average daily outbound volume during the campaign 800 units, remaining stock 5000 units." The draft sentence to be judged is "This product's campaign price is 29.9 yuan, stock is available throughout the campaign, and orders ship the same day."
Configuration and implementation checks
| Configuration item | Value or check method on this page | Basis and explanation |
|---|---|---|
| Linked knowledge bases | Select the registration form, cost sheet, refund record, and in-site message bases that use the same vector model | The node can link one or more bases sharing the same vector model |
| Search mode | Compare semantic, full-text, and hybrid retrieval for "what is the campaign price" and "what is the registration deadline" | Compare per test question; no single mode is assumed |
| Citation limit | For this hypothetical exercise, start at 2000 tokens and compare completeness against actual excerpts of differing length | Unit is tokens; adjust after testing with real materials |
| Minimum relevance | Label relevant and irrelevant samples first, then set a hypothetical exercise starting value of 0.5 | Takes effect only under semantic retrieval or result reranking |
| Question optimization | Compare the original and optimized forms of "how much is this" and check whether the subject was altered | Used for coreference resolution and question expansion |
| Result reranking | Have a rerank model available and enabled by configuration to reorder candidate text | Reranked score is 0-1 and can be used for filtering |
Three common errors
- Forcing knowledge bases with different vector models into one link. Input is two bases with different vector model identifiers; compare whether the identifiers match, exclude those that do not, and record the actual observation. Do not run a cross-model search to compare missed recalls.
- Treating minimum relevance as a permission or entity filter. Input is a similarity threshold setting; compare whether returned content includes records that are not authorized. A relevance score does not perform permission filtering, and retrieval itself is not authorization.
- Treating "stock available throughout the campaign" and "ships the same day" in the draft sentence as supported by the materials. Input is the remaining stock and projected outbound volume in the cost sheet; compare whether the draft sentence contains commitments the materials do not state. Content not stated remains unsupported and the evidence gap is preserved.
How to verify the setup
- Input the source material "The registration deadline for this campaign is 12 March at 18:00" together with the draft sentence "This product's campaign price is 29.9 yuan, stock is available throughout the campaign, and orders ship the same day." Compare whether the draft sentence contains supply and shipping commitments not stated in the materials. If the draft sentence includes "stock available throughout the campaign" or "ships the same day" while the materials do not state them, the claim is judged unsupported.
- Input the draft sentence "This product's campaign price is 29.9 yuan" together with the cost sheet text "campaign price 29.9 yuan." Compare whether the price value and the product subject match. If both match, the price sentence has material support; if the subject or the value differs, the item remains pending verification.
- Input an empty citation array [] together with the downstream node. Compare whether the downstream node continues to run with array length 0. If the downstream node accepts the empty result normally, the path design holds; if the downstream node errors, record the actual phenomenon and do not infer that a fictional rule must return an empty array.