What the documents look like
Pinduoduo low-price campaign operations revolve around campaign resource slots, campaign inventory, and refund-rate fluctuations. The working materials are campaign registration forms, product cost calculation sheets, after-sales refund orders, and platform internal messages. These materials update daily and per campaign. In store service and internal training, staff must place these materials into a knowledge base so that campaign operators and supply-chain buyers can retrieve them. The materials mix structured fields with narrative passages, so retrieval-word quality and data-chunk quality jointly determine recall.
For this page, the relevant task is knowledge-base retrieval and recall. The knowledge-base search node can associate one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. When identifiers differ, the correct action is to check the identifier, exclude the non-matching base, and record what was actually observed. Semantic retrieval, full-text retrieval, and hybrid retrieval suit different material types. Numbers and proper nouns fit full-text or hybrid retrieval. The minimum relevance setting takes effect only in semantic retrieval or when result reranking is used. The citation limit is measured in tokens. Question optimization performs coreference resolution and question expansion, and it adds an extra model call. Citation content is output as an array, and its length can be zero.
Constraints at the retrieval step
Retrieval returns candidate content. It does not authorize a business action. A refund order with status pending review is not a completed refund. A search hit on a registration form is not proof that a campaign slot was granted. Staff training must keep these stages separate: what the document states, what the node returns, what a human verifies, and what a downstream system must implement.
A hypothetical paper exercise sample is given here for checking. A store training knowledge base contains this excerpt from a campaign registration form: "The registration deadline for this campaign is March 12 at 18:00. The registered product is item A. The campaign inventory ceiling is 500 units. The supply price follows the current-day version of the cost calculation sheet." A separate after-sales refund order reads: "Order number H-0001. Refund application submitted March 15. Status is pending review." The draft sentence to be judged is: "This order has completed the refund and can be counted into this campaign's refund rate." The source text states only the pending-review status. It does not state a review outcome or a completed refund. The draft sentence is not supported by the source.
Question optimization is useful for multi-turn follow-ups and short questions. It adds a model call and therefore increases latency. If the current question is already clear, or if response speed matters more, enabling it is a practical decision rather than a fixed rule. Reranking requires a rerank model and must be enabled under the settings. Image retrieval requires an image input and the relevant model capability. The optional paths in the overall retrieval flow are not all executed on every request.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Associate only the store training base that uses the same vector model | Check each base's vector model identifier; exclude non-matching ones |
| Search mode | Test semantic retrieval first, then compare with full-text retrieval results | Use the same set of test questions to compare recalled content across modes |
| Citation limit | Start at a token amount that covers one complete campaign registration passage | Compare against actual materials of different lengths to see whether content stays complete |
| Minimum relevance | Start at 0.5, effective only in semantic retrieval or reranking | Label relevant and irrelevant samples first, then compare against the threshold |
| Question optimization | Enable first in multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous | Requires a rerank model enabled under settings; compare ordering before and after reranking |
Three common errors
- Treating a retrieval hit as authorization to execute. Input: the pending-review refund order. Comparison object: the draft sentence claiming a completed refund. Observable result: the source status field and the draft sentence disagree. Retrieving that order does not mean the refund was executed.
- Associating knowledge bases across different vector models directly. Input: two bases whose vector model identifiers differ. Comparison object: the node's condition for permitted association. Observable result: the non-matching base is excluded and the observation is recorded. Do not continue by inventing a runtime outcome.
- Treating minimum relevance as a permission or entity filter. Input: an unrelated business question. Comparison object: the returned citation content. Observable result: low-relevance content is filtered by the threshold. Authorization scope still requires downstream implementation checks.
How to verify the setup
- Input the campaign registration form excerpt and the draft sentence from the hypothetical sample. Compare the deadline, product, and inventory ceiling stated in the source with the wording of the draft sentence. Observe whether the draft sentence matches the source. A mismatch means the draft sentence is not supported.
- Input the pending-review refund order from the hypothetical sample. Compare the source status field with the wording "completed the refund." Observe whether the two are mutually exclusive. When the status is pending review, the draft sentence remains to be verified.
- Input a clearly paper-based empty result
[]. Compare how the downstream node is designed to handle an empty citation. Observe whether the chain still runs through. Do not claim that a fictional rule search will necessarily return[].
These checks test the named primary task: retrieval and recall for staff training. They do not test identity certification, report analysis, or any other task. Where evidence is missing, it remains missing. A pending request is not approval. An offered option is not execution. An unsigned termination is not a completed termination.