What the documents look like
Pinduoduo low-price campaign operations revolve around low-price resource slots, campaign inventory and refund-rate fluctuation. The daily materials named in the brief are the campaign registration form, the product cost calculation sheet, the after-sales refund slip and platform site messages. At a handover, these materials sit at different stages. A registration form may be submitted and awaiting review. A cost calculation sheet may be a draft. A refund slip may be accepted but not yet closed. A site message may be a notice only. Knowledge base search and recall lets the incoming operator find historical records among materials already ingested. The retrieval result itself does not change the review, signature or execution state of any material.
Constraints at the retrieval step
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. The vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded. The output is an array of quoted content whose length can be 0, so the downstream path still runs when nothing is found. Minimum relevance takes effect only in semantic search or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The quote cap is measured in tokens and controls how much content one search may quote. Question optimization performs coreference resolution and question expansion and adds one model call. Hybrid search merges multiple retrieval paths through RRF; RRF is a fusion process, not a separate switch that can be turned on. Image retrieval requires image input and the corresponding model capability, and reranking requires a model and must be enabled per settings. These optional paths do not mean every request executes all of them.
Hypothetical paper exercise sample: a Pinduoduo campaign operator is handing over after leaving, and three materials have been ingested into a knowledge base. Material one is a campaign registration form whose source text reads "Campaign registration form: Store A, campaign slot March 2025 low-price resource position, submission date 20 February 2025, status is submitted, awaiting review." Material two is a product cost calculation sheet whose source text reads "Product cost calculation sheet: Store A, product code P001, calculation date 18 February 2025, status: draft, not submitted for approval." Material three is an after-sales refund slip whose source text reads "After-sales refund slip: Store A, order number 20250215001, refund application date 15 February 2025, status: accepted, not closed." The draft sentence to be judged reads "Store A's March 2025 low-price resource position campaign registration has been approved, the cost calculation has passed approval, and the refund slip has been closed." The check is whether the retrieved source text and the draft sentence share the same subject, the same period and the same field.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge base | Associate one knowledge base containing the three materials above | Check that the knowledge base vector model identifier matches; bases on different models are not associated together |
| Search mode | Test semantic search first, then compare with full-text search | Compare quoted content for the same question under both modes and record the actual observation |
| Quote cap | Hypothetical starting point of 2000 tokens | Chosen because the three materials are short in total; compare content completeness against a 500-token result |
| Minimum relevance | Hypothetical starting point of 0.3 | Chosen to filter low-relevance fragments; compare quoted content changes against 0.1 and 0.5 |
| Question optimization | Off first, then compare the same question with it on | Compare whether the subject in the original question and the optimized question has been altered incorrectly |
| Result reranking | Not enabled first; enable per settings once a model is available | Compare quote ordering before and after enabling and record the actual observation |
Three common errors
- Treating a retrieved registration form marked "submitted, awaiting review" as "approved". The error is taking an awaiting-review status as completed approval. Check method: input the draft sentence and the source text of material one, compare the status field values, and if the source reads "submitted, awaiting review" while the draft sentence says "approved", the draft sentence is unsupported on that field.
- Treating a cost calculation sheet marked "draft" as "passed approval". The error is taking a draft status as approval completion. Check method: input the draft sentence and the source text of material two, compare the status field values, and if the source reads "draft, not submitted for approval" while the draft sentence says "passed approval", the draft sentence is unsupported on that field.
- Treating a refund slip marked "accepted, not closed" as "closed". The error is taking an acceptance status as closure. Check method: input the draft sentence and the source text of material three, compare the status field values, and if the source reads "accepted, not closed" while the draft sentence says "closed", the draft sentence is unsupported on that field.
How to verify the setup
- Input the source text of material one and the draft sentence, compare the status field "submitted, awaiting review" against "approved", and observe whether both share the same subject, the same period and the same field with mutually exclusive values. If the values are mutually exclusive, the draft sentence is unsupported on that field.
- Input the source text of material two and the draft sentence, compare the status field "draft, not submitted for approval" against "passed approval", and observe whether both share the same subject, the same period and the same field with mutually exclusive values. If the values are mutually exclusive, the draft sentence is unsupported on that field.
- Input the source text of material three and the draft sentence, compare the status field "accepted, not closed" against "closed", and observe whether both share the same subject, the same period and the same field with mutually exclusive values. If the values are mutually exclusive, the draft sentence is unsupported on that field.
All three verifications use paper materials as input and do not claim that retrieval or testing was actually executed.