What the documents look like
Pinduoduo low-price campaign operations revolve around low-price 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 internal messages. Data sources named in the brief are the Pinduoduo merchant backend, campaign resource-slot data, ERP inventory and logistics receipt data, with updates described as daily or per campaign. The three most time-consuming steps are estimating campaign inventory, tracking abnormal refunds, and checking gross margin at the campaign price.
For operating reviews and exception reviews, retrieval is asked to bring back the right passages from those four document types. A registration form states which product was submitted, at what campaign price, in what quantity, and at what registration status. A cost calculation sheet states per-unit cost and the date of calculation. A refund record states refund quantity, reason and refund status. An internal message states what the platform wrote. These are different business facts. Retrieval returns quoted content only; it does not grant authorization, approve anything or write to any system.
Constraints at the retrieval step
The knowledge base search node can be linked to one or more knowledge bases that use the same vector model. Semantic search, full-text search and hybrid search each have their own conditions. Hybrid search merges multiple retrieval channels through RRF, and ReRank performs a second ordering of candidate text. The reference limit is measured in tokens. The minimum relevance value takes effect only under semantic search or when result reordering is used. Question optimization performs coreference resolution and question expansion. Image retrieval requires an image input and matching model capability; reordering requires a model and is enabled according to settings. Retrieval output is not an authorization decision, and a similarity score does not perform entity isolation or permission filtering.
Hypothetical paper exercise. A campaign operator intends to check one low-price campaign's registration and cost materials. The supplied source text reads: "The campaign registration form states that the registered product is Style A, the registered campaign price is 9.9 yuan, the registered quantity is 500 units, and the registration status is submitted." "The product cost calculation sheet states that the per-unit cost of Style A is 7.2 yuan, including packaging and logistics allocation, and the calculation date is the registration date." "The after-sales refund record states that 12 units of Style A were refunded during the campaign, the refund reason is incorrect size, and the refund status is completed." The draft sentence to be judged reads: "Style A was registered at a campaign price of 9.9 yuan for 500 units, the per-unit cost is 7.2 yuan, and 12 units were refunded during the campaign with the refund completed." Judging this draft requires source support on the same subject, the same period and the same field. Authorization status, approval outcome and inventory deduction are not stated in the materials and remain unknown.
Configuration and implementation checks
| Configuration item | Assumed starting point for this page | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Link one base containing the registration form and the cost calculation sheet | Check that this base and the search node use the same vector model; exclude bases whose model identifier differs |
| Search mode | Semantic search combined with full-text search | Test with a registration form number and a product name separately; compare whether the number-type question reaches the full-text path |
| Reference limit | Set to 2000 tokens as the assumed trial starting point | Compare returned content completeness using long and short registration and refund passages, then adjust |
| Minimum relevance | Set to 0.3 as the assumed trial starting point | Label relevant and irrelevant samples first; compare whether registration subject and amount fields survive filtering |
| Question optimization | Enabled for multi-turn follow-up scenarios | Compare the original question with the optimized question; check whether the campaign name and product subject were altered |
| Result reordering | Enabled when candidates are numerous | Confirm a reordering model is configured; compare reordering scores against original scores for the same candidate |
Three common errors
- Treating a retrieval score as an authorization or entity-isolation basis. Input is a refund-record retrieval request; the comparison item is the similarity score against the status field in the refund record source text; the observable result is that a high or low score does not change the completed or pending-review status of the refund record, and the score does not perform permission filtering.
- Describing the RRF merge of hybrid search as an independent toggle that can be switched on. Input is the search parameter configuration; the comparison item is the search modes and RRF description listed in the documentation; the observable result is that RRF is a merge process, and no configuration item named RRF exists.
- Treating an unstated authorization status as already approved. Input is the registration-status wording in the draft sentence; the comparison item is the registration form source text, which states only submitted; the observable result is that submitted does not equal reviewed and approved, and the authorization status remains unknown.
How to verify the setup
- Input the draft sentence from the hypothetical exercise together with the three source texts. Compare whether the campaign price, registered quantity, per-unit cost and refund quantity in the draft sentence match the source texts on the same subject, the same period and the same field. The condition is met when all four values have a clear origin in the corresponding material and no authorization or approval conclusion absent from the source has been added.
- Input a query containing only a registration form number. Compare the returned content from semantic search against full-text search. The condition is met when the number-type exact query reaches the registration form passage on the full-text path, and the semantic path's returned content does not replace source verification.
- Input a clearly paper-based empty result []. Compare how the downstream design receives an empty reference. The condition is met when the output chain still completes with a reference content length of zero, and the downstream does not interpret the empty result as no risk or no exception. All of the above are paper verification expectations; no actual system test was executed.