What the documents look like
Pinduoduo low-price campaign operations revolve around low-price resource slots, campaign inventory and refund-rate swings. Materials update daily or once per campaign. This page covers four document types: campaign registration forms, product cost calculation sheets, after-sales refund slips and platform in-site messages. For inventory procurement and supplier handover, the decisive facts sit in three of them. A campaign registration form states the registered product, the campaign price and the campaign inventory quantity. A product cost calculation sheet states procurement cost and gross-margin calculation. An after-sales refund slip records the refund reason and amount. Together these materials determine whether a procurement or supplier handover request has checkable grounds. Knowledge-base retrieval here only recalls these materials by semantics or keyword for human comparison. Retrieval itself does not verify authorization status.
Constraints at the retrieval step
The knowledge-base search node can associate one or more libraries that use the same vector model. Semantic retrieval depends on vector distance, full-text retrieval depends on keyword hits, hybrid retrieval merges the two paths through RRF, and result reranking uses a ReRank model for a second ordering. Minimum relevance takes effect only under semantic retrieval or when result reranking is used. The reference cap is measured in tokens. Question optimization performs coreference resolution and question expansion, and adds one model call.
Hypothetical drill sample: the knowledge base contains the original text of a Campaign Registration Form (hypothetical number HD-001) reading "Campaign product: A-type storage box; campaign price: 9.9 yuan; campaign inventory: 500 units; registration status is pending review." A Product Cost Calculation Sheet (hypothetical number CB-001) reads "A-type storage box procurement cost 7.5 yuan; logistics cost 1.2 yuan; calculated gross margin 1.2 yuan." The handover request to be judged is: "HD-001 has passed review, place an order with the supplier for 500 units." For checking, retrieve "HD-001 registration status" and compare whether the recalled original text states "passed review." The original text states only "pending review," so the request is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this page | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge base | Associate only the library holding the campaign registration form and the cost calculation sheet | First check each library's vector model identifier; libraries with mismatched identifiers are not associated together |
| Search mode | Test semantic retrieval and full-text retrieval once each | Use an identifier such as "HD-001" for full-text, and "how to estimate campaign inventory" for semantic |
| Reference cap | Hypothetical starting point of 2000 tokens | Starting point for this trial; compare whether recalled content is complete using registration-form fragments of different lengths |
| Minimum relevance | Hypothetical starting point of 0.4 | Starting point for this trial; label relevant and irrelevant samples first, then compare threshold changes |
| Question optimization | Off first, testing with complete questions | Compare the original question with the optimized question and check whether the subject was altered incorrectly |
| Result reranking | Not enabled first | Confirm a rerank model is available and enabled by configuration before comparing ordering |
These values are hypothetical trial starting points, not official defaults.
Three common errors
- Treating retrieval recall as grounds for authorization. Input the request "HD-001 has passed review"; retrieval may recall the registration form original text, but its status is "pending review." The error is substituting recall results for a status check. Check: input that request, compare the status field in the recalled original text, and observe whether the wording "passed review" appears; if it does not appear, record the request as unsupported.
- Associating libraries that use different vector models. Input the model identifiers of two libraries and compare whether they match. The error is skipping the identifier check and associating directly. Check: read the vector model identifier of each library, exclude those with mismatched identifiers, and record the actual observation without asserting runtime behavior after association.
- Using minimum relevance filtering as a substitute for entity isolation. Input an unrelated business query and compare whether recalled content belongs to the target campaign. The error is assuming a higher relevance value excludes unrelated records. Check: observe the correspondence between recalled content and the query subject; whether unrelated records are excluded remains unsupported and requires manual review.
How to verify the setup
- For hypothetical sample HD-001, input the search term "HD-001 registration status" and compare the recalled original text with the drafted sentence "passed review." Expected observation: the recalled original text shows "pending review," the drafted sentence does not match the original text's status field, and the judgment is unsupported.
- For hypothetical sample CB-001, input the search term "A-type storage box procurement cost" and compare the recalled amount with the figure in the calculation sheet. Expected observation: the recalled original text shows 7.5 yuan, consistent with the calculation sheet, and the judgment is checkable; if the recall is an empty array, handle it as an empty result against the downstream design.
- For cross-library association, input the vector model identifiers of two libraries and compare whether the identifiers match. Expected observation: only matching identifiers may be associated together; mismatched ones are excluded and recorded, and no cross-model retrieval is performed to compare missed recall.
All three verifications use paper hypothetical samples and were not deployed or executed.