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 are campaign registration forms, product cost calculation sheets, after-sales refund slips and platform in-site messages. Registration forms and cost sheets carry campaign price, cost basis and registration conditions. Refund slips record the transaction-side refund state. Platform in-site messages carry platform notices and requirements. These materials update daily or once per campaign.
For contract and transaction material comparison, the relevant unit is one campaign, one product and one period. A registration form often places the condition and the product, price and period it constrains inside the same passage. A cost sheet places cost items next to their calculation basis. A refund slip places refund state next to the order identifier. If chunking separates a condition from what it constrains, the retrieved fragment loses the semantics needed for a judgement. One group of source data can map to several indexes, which allows one material to be indexed by product, by campaign and by period.
Constraints at the document chunking step
Comparison needs evidence from the same subject, the same period and the same field. Chunking must keep subject, object, state or validity range inside one data group. Retrieval uses semantic search, full-text search or hybrid search to recall candidate content. Hybrid search merges multiple result paths through RRF, and reranking performs a second ordering of candidate text. Similarity and relevance affect recall and ordering only; they do not perform permission isolation and do not decide whether a record belongs to the current shop or the current campaign.
Full-text search suits exact queries such as numbers, models and proper nouns. Semantic search suits natural-language phrasing. Campaign numbers, product numbers and order identifiers in these materials suit full-text checking; descriptive content such as cost basis and refund reason suits semantic search. The citation limit is controlled in tokens. The minimum relevance value takes effect only in semantic search or when result reranking is used. Image retrieval requires corresponding image input and model capability, and reranking requires a model and must be enabled in settings. Optional paths shown in the document flow do not mean every request executes all of them.
Configuration and implementation checks
The following configuration is listed around a hypothetical drill sample. The values are starting points for this hypothetical trial, used to explain the reason for the choice and the comparison object. They are not official default values.
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Search mode | Hybrid search | The sample contains campaign numbers and descriptive refund reasons, so both exact and semantic recall are needed; compare with semantic search alone |
| Minimum relevance | 0.5 | The sample question is explicit, so the starting point takes a middle value; compare filtered candidate counts at 0.3 and 0.7 |
| Citation limit | 3000 tokens | Sample passages are short, so the starting point covers several candidates; compare citation completeness at 1500 and 6000 |
| Question optimisation | Not enabled for now | The sample question has a complete subject, so observe recall without optimisation first; compare with recall after enabling |
| Result reranking | Not enabled for now | The sample is expected to yield few candidates, so observe ordering without reranking first; compare with ranking positions after enabling |
| Chunk boundary | By material passage | In the sample the condition and what it constrains share a passage, so split by passage; compare passage completeness with fixed-length splitting |
Three common errors
- Splitting a condition and what it constrains into different data groups. In the hypothetical drill sample, the campaign registration form reads: "The product registered for this campaign is Model A, the campaign price is 9.9 yuan, and the campaign period is 1 June to 3 June." If chunking moves "the campaign period is 1 June to 3 June" into another group, the retrieved fragment lacks the period, and comparison cannot determine whether the refund slip falls inside the campaign period.
- Using a similarity score in place of subject and period checking. In the hypothetical sample, the product cost calculation sheet reads: "Model A cost 7.5 yuan, including packaging and logistics." After this passage is retrieved, treating a high score as proof that it belongs to the same campaign as the refund slip ignores whether the two materials point to the same campaign number and the same period.
- Treating retrieved content as authorisation or as an executed state. In the hypothetical sample, the platform in-site message reads: "Please confirm Model A campaign inventory before 1 June." Retrieving this sentence shows only that the notice exists. It does not show that inventory has been confirmed, that the campaign has been registered, or that a contract has taken effect.
How to verify the setup
The following verification uses the same hypothetical drill sample. The sample material text is: campaign registration form, "The product registered for this campaign is Model A, the campaign price is 9.9 yuan, and the campaign period is 1 June to 3 June."; product cost calculation sheet, "Model A cost 7.5 yuan, including packaging and logistics."; after-sales refund slip, "Order 20240602-001, Model A, refund state: pending review."; platform in-site message, "Please confirm Model A campaign inventory before 1 June." The draft sentence to be judged is: "Model A had a campaign price of 9.9 yuan during 1 June to 3 June, and the after-sales refund slip has passed review."
- Input the campaign registration form text together with the first half of the draft sentence. Compare the campaign price of 9.9 yuan and the campaign period of 1 June to 3 June against the source text, and observe whether the draft sentence fully preserves the source subject, object and period. If the period is rewritten or omitted, the sentence does not meet the condition.
- Input the after-sales refund slip text together with the second half of the draft sentence. Compare the source "refund state: pending review" against the draft "has passed review", and observe whether the two give mutually exclusive values for the same field of the same order. The source states a pending-review state and the draft writes passed review, which is a same-field contradiction, so the sentence does not meet the condition.
- Input the platform in-site message text together with a handover request, "Model A campaign inventory is confirmed, registration can proceed directly". Compare the source "please confirm before 1 June" against "is confirmed" in the request, and observe whether the request turns a pending task notice into a completed fact. The source records only a confirmation requirement and records no confirmation result, so the request is unsupported; record the evidence gap and keep the state as pending verification.