What the documents look like
Pinduoduo low-price campaign operations revolve around campaign resource slots, campaign inventory, and refund-rate swings. Product information and selection questions draw on four document types: campaign registration forms, product cost calculation sheets, after-sales refund records, and platform internal messages. These documents update daily or per campaign, and the same product can carry different registration prices, inventory commitments, and refund records across campaign sessions.
Citation tracing in this setting means that any answer about whether a product can join a campaign, or at what price it can be supplied, must point back to a specific document and a specific field. Historical impressions or inferences from similar products are not a source.
The knowledge base search node can select one or more knowledge bases that share the same vector model, and it outputs cited content as an array whose length may be zero. Even when nothing is retrieved, the output path still runs. Search mode, result reranking, and question optimization are optional paths. Image retrieval requires corresponding image input and model capability, and reranking requires a rerank model that is enabled by configuration. These paths do not all execute on every request. Minimum relevance 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 constrains knowledge base association only and does not extend to external systems. RRF is the fusion process for mixed results, not a switch that can be turned on.
Constraints at the citation tracing step
The citation array passed to the answering step must be checked item by item against the source text. A registration form that records a status of pending review does not show that review has passed. A cost calculation sheet that records unit cost and packaging and logistics totals does not record a supply deadline or a committed supply quantity. A retrieval result is not an authorization to submit a registration or lock inventory.
Hypothetical drill sample: the source text reads, "The campaign registration form submitted by Shop A on March 8, 2025 for a certain insulated cup records: campaign price 39 yuan, registered inventory 500 units, registration status pending review." A second document, a product cost calculation sheet dated the same day, reads, "Unit cost for this insulated cup is 34 yuan, packaging and logistics total 3 yuan." The draft sentence to be judged is: "Shop A's insulated cup has passed the March 8 campaign registration and can supply 500 units at 39 yuan." In the check, the registration form shows only pending review, not approved review; the cost sheet does not state a supply deadline. The phrases "has passed registration" and "can supply 500 units at 39 yuan" both lack source support.
Configuration and implementation checks
| Configuration or check item | Use on this page | Basis and operating boundary |
|---|---|---|
| Associated knowledge bases | Limit the retrieval scope for product information and selection questions | Select only knowledge bases sharing the same vector model; verify identifiers first and exclude bases with different vector models |
| Cited content output | Carry retrieval results to the answering step | Output is an array whose length may be zero; an empty array does not mean the workflow terminated with an error |
| Search mode | Determine semantic, full-text, or hybrid recall paths | Choose per the documented options; hybrid retrieval is fused through RRF and is not written as an on/off switch |
| Result reranking | Reorder candidate text a second time | Requires a rerank model enabled by configuration; the rerank score is 0-1 and can be used for filtering |
| Minimum relevance | Filter low-relevance results | Takes effect only in semantic retrieval or when result reranking is used; for this hypothetical trial the starting value is set to 0.3 because the sample question contains a specific product name and date, and the comparison is against the number of recalled items with no threshold set |
| Citation limit | Control the tokens of content cited per search | Measured in tokens; for this hypothetical trial the starting value is set to 2000 because the sample materials are short, and the comparison is against citation completeness at 1000 |
Three common errors
- Treating pending review as passed. The registration form records the registration status as pending review, and the draft sentence says "has passed registration." The input is the registration form source text and the draft sentence, the comparison object is the registration status field, and the observable result is that the draft sentence does not match the source status.
- Treating cost calculation as a supply commitment. The cost calculation sheet records only unit cost and the packaging and logistics total, with no supply deadline or available quantity. The draft sentence says "can supply 500 units at 39 yuan." The input is the cost sheet source text and the draft sentence, the comparison object is the supply quantity and deadline fields, and the observable result is that the source does not record those fields.
- Treating retrieval results as authorized execution. The search node outputs a citation array whose length may be zero. If the citation array is treated directly as evidence that registration was submitted or inventory was locked, the input is the citation array and the handoff request, the comparison object is the authorization status field, and the observable result is that the citation array contains no authorization field.
How to verify the setup
- Input the registration form source text and the draft sentence from the hypothetical drill sample, compare the draft sentence's registration status wording against the source's "pending review," and observe that the draft sentence is marked as unsupported while the pending-review status remains to be verified.
- Input the cost calculation sheet source text and the draft sentence from the hypothetical drill sample, compare the draft sentence's supply quantity and deadline wording against the fields recorded in the source, and observe that the source does not record a supply deadline or available quantity, so that part is judged an evidence gap.
- Input an explicit paper citation array of [], compare whether the downstream answering step accepts the empty citation as designed, and observe that the output path runs through without producing source citations; this result does not mean that a retrieval of fabricated rules will necessarily return an empty array.