What the documents look like
Pinduoduo low-price campaign store operations revolve around campaign resource slots, campaign inventory and refund-rate swings. Four document families form the daily working set: campaign registration forms, product cost worksheets, after-sales refund records and platform in-site messages. These update daily or once per campaign. The forms differ in shape. A registration form carries a campaign name, the enrolled product, the campaign price and the enrolled quantity. A cost worksheet carries cost line items and the calculation basis. A refund record carries an order, a refund reason and a handling status. An in-site message carries a platform notice and a deadline. Retrieval and recall operate over this heterogeneous text. The search unit is the library; a collection only classifies data and does not affect search results. Fields not stated in a document remain unknown.
Constraints at the retrieval step
Search runs against a library. A single search node can associate one or more knowledge bases 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. The minimum relevance value applies only under semantic retrieval or when result reranking is used, and it acts as a filter threshold rather than a permission or entity boundary. The citation limit is measured in tokens. Citation output is an array whose length can be 0, and the downstream path still runs when nothing is retrieved. Retrieval on its own does not constitute authorization, and it does not mean that business approval, a database write or an external synchronization has occurred.
Hypothetical drill sample. A store operator prepares a campaign registration record. The source text reads: "Campaign name: Summer clearance; enrolled product: Model A storage box; campaign price: 9.9 yuan; enrolled quantity: 500 units; registration deadline: July 20." A separate cost worksheet states: "Model A storage box unit cost 8.5 yuan, including packaging and logistics." The draft sentence to be judged is: "Model A storage box has passed campaign registration review and can be listed at 9.9 yuan." That sentence concerns a review status. The source text states no review conclusion, so the claim is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this drill | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the single vector-model library holding the registration form and cost worksheet | Check each library's vector model identifier first, exclude mismatches, and do not run cross-model retrieval |
| Search mode | Begin with hybrid retrieval | The registration form contains natural language such as campaign names and also product-code fields; compare the semantic and full-text result sets |
| Citation limit | Hypothetical starting point of 2000 tokens | Test with registration material of differing lengths to see whether content is truncated, and compare against a shorter value for completeness |
| Minimum relevance | Hypothetical starting point of 0.4 | Label relevant and irrelevant samples first, observe how filtered citation content changes, then adjust |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when a rerank model is available | Compare scores and ordering before and after reranking, and filter candidates by score |
These values are hypothetical starting points for this drill, not official defaults.
Three common errors
- Treating a retrieval hit as business authorization. The input is the draft sentence "Model A storage box has passed campaign registration review." Retrieval may return the registration source text, but that text states only registration details and no review conclusion. Compare the fields in the retrieval result against the assertion in the draft sentence; the missing review status means the claim is unsupported.
- Associating libraries that use different vector models together. The input is two libraries with different vector model identifiers. Compare whether the identifiers match; exclude the mismatched one and record it, and do not run cross-model retrieval to compare missed recalls.
- Treating minimum relevance as a permission filter. The input is an unrelated business query. Compare citation content before and after filtering; the relevance threshold only affects low-relevance results and does not exclude unauthorized records.
How to verify the setup
- Input the hypothetical source text and the draft sentence, then compare the assertion in the draft sentence against the fields stated in the source. The source contains campaign name, product, campaign price, quantity and deadline, and contains no review conclusion. The draft sentence asserts review passage, so the judgment is unsupported.
- Input an empty-citation scenario, using the explicit paper input
[]to check the downstream path design, and confirm that the output path still runs when citation length is 0. Do not claim that a fictional rule search necessarily returns an empty array. - Input a query containing a product code and compare the recall difference between semantic retrieval and full-text retrieval, observing whether the code-type field is hit by the full-text path. If it is not hit, record that observation and do not infer a deterministic failure cause.
A pending review status does not demonstrate a lack of operating permission, and a retrieval result does not constitute approval, signing or contract effectiveness. Missing evidence remains missing; record the gap.