What the documents look like
Group-buy presale collection takes orders first and organizes procurement and fulfillment afterward. Copy drafting for products and marketing depends on the presale product table, the collection result, the procurement plan and the cut-off notice. These documents update daily or per collection session. Claims about specifications, timing and group-formation conditions in a draft sentence need to match the corresponding document fields.
The central risk at this step is whether a proposed promotional sentence has same-subject, same-period, same-field textual support in those documents. Citation tracing requires checking each draft sentence against the source text item by item. When the source is missing, the gap stays open. A search result or a model answer does not substitute for source evidence.
Two industry facts shape this task. First, the sequence is order collection before procurement and fulfillment, so a draft sentence that asserts fulfillment timing depends on documents that may not yet contain that commitment. Second, the daily or per-session update cycle means the version of a document matters when a sentence is checked against it.
Constraints at the citation tracing step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Its citation output is an array, and the length can be zero. An empty array does not mean the workflow has errored and stopped. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used; it is a filtering threshold and does not perform permission or entity isolation. Hybrid retrieval merges multiple result paths through RRF. A rerank model reorders candidate text. Image retrieval needs corresponding image input and model capability, and reranking needs a model and must be enabled in settings. These optional paths do not mean every request executes all of them. The citation limit controls the maximum content cited per search in tokens, and it is unrelated to thinking budget.
Hypothetical paper exercise: on 10 June 2025, a platform operator checks one draft sentence for a presale item. The source text provided for checking has three sentences. First, the presale product table states: "Red-fleshed dragon fruit, specification 5-jin pack, cut-off time 12 June 20:00." Second, the collection result states: "As of 11 June 18:00, the collected quantity for this item is 120 units, below the 200-unit dispatch threshold." Third, the cut-off notice states: "If the group-formation threshold is not reached at cut-off, the order is automatically refunded." The draft sentence to be judged is: "Red-fleshed dragon fruit 5-jin pack, group formed, delivered the day after cut-off." The check compares the subject, object, status and effective scope of the draft sentence against the three source sentences.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Associated knowledge bases | One base each for the presale product table, the collection result and the cut-off notice | The three bases must use the same vector model to be associated together; check the vector model identifier one by one and exclude any that does not match |
| Search mode | Hybrid retrieval | Exact fields such as specification and time need full-text recall, and semantically similar wording needs vector recall; the two paths merge through RRF |
| Minimum relevance | Hypothetical starting value 0.3 | This is a 0-1 value that takes effect only in semantic retrieval or result reranking; compare with 0.5 to observe changes in recall range |
| Citation limit | Hypothetical starting value 2000 tokens | Measured in tokens; compare whether different values cover the key fields of the three documents in the cited content |
| Result reranking | Enabled | When candidate results are numerous, text is reordered a second time and the rerank score is used for filtering; compare ordering before and after enabling |
| Question optimization | Enabled | When the draft sentence contains references or omissions, the search terms are completed; compare whether cited content contains the target fields before and after enabling |
Three common errors
- Treating minimum relevance as a permission or entity filter. The input is the draft sentence and the three knowledge bases; the comparison object is the minimum relevance setting and the cited content returned by retrieval; the observation is whether records for other products unrelated to the draft sentence subject appear in the results. Relevance filtering does not exclude unrelated business records. When unrelated records appear, record the actual observation and do not infer a cause.
- Judging an empty citation array as a workflow failure. The input is one retrieval request; the comparison object is the node output citation array and an empty array; the observation is whether the downstream answer step can still proceed. When the output length is zero, the path can still continue, and the downstream design for accepting empty input needs to be checked.
- Treating a retrieval hit as support for the copy. The input is the draft sentence "group formed, delivered the day after cut-off"; the comparison objects are the collection result text "below the dispatch threshold" and the cut-off notice text "if the group-formation threshold is not reached, the order is automatically refunded"; the observation is whether the status in the draft sentence matches the source text in subject, period and field. When the source text does not state next-day delivery, that statement is unsupported and the evidence gap stays open.
How to verify the setup
- Input the three source sentences and the draft sentence from the hypothetical exercise. Compare each assertion in the draft sentence with the subject, object, status and effective scope in the source text, and observe whether a mutually exclusive value appears in the same field. The draft sentence says "group formed," while the collection result says "below the dispatch threshold." These are contradictory values in the same field and should be judged unsupported.
- Input a retrieval request containing only "red-fleshed dragon fruit." Compare the returned cited content with the three source documents and observe whether each cited item can be traced back to the source text. Cited content needs character-by-character comparison with the source, and answer fluency does not substitute for source evidence.
- Input a paper empty array [] as the retrieval output. Compare the downstream answer step's design for accepting it and observe whether empty input is handled as a branch with no citation available. This check verifies the downstream design and does not claim that a fictional rule retrieval necessarily returns an empty array.