What the documents look like
Group-buy presale orders collect commitments first and organize procurement and fulfillment afterward. The comparison task in this category works across four material types: presale product tables, group-order results, procurement plans, and cut-off notices. These materials update by day or by session, and field definitions can shift between the pre-cut-off and post-cut-off versions of the same record. The comparison question is narrow: for one subject, one period, and one field, do the values agree? Retrieval returns candidate source text. It does not decide authorization status, and it does not replace a human confirmation step.
Because a presale session produces a product table before orders close and a group-order result after orders close, the same product line can appear in two documents with different quantities at different stages. A procurement plan then restates a quantity and a supply window. A cut-off notice fixes the time boundary. A comparison workflow that retrieves all four types together can surface these records in one pass, but the operator still has to check whether the subject, period, and field match before calling anything consistent or contradictory.
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 using different vector models cannot be associated together. Retrieval output is an array, and its length can be zero; the downstream path still runs when nothing is retrieved. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. The citation limit is measured in tokens. RRF is a fusion process for combining result paths, not an on/off switch. Image retrieval requires an image input and model capability. Reranking requires a configured model and must be enabled in settings. Retrieval and similarity scores do not perform permission or entity isolation.
Hypothetical paper exercise sample: the source text to be checked reads, "Presale product table A-0712 states: cut-off time 12 July 2025 at 18:00; product is a leafy vegetable assortment; planned quantity 300 units. Group-order result A-0712 states: same cut-off time; leafy vegetable assortment group-order quantity 280 units. Procurement plan P-0712 states: leafy vegetable assortment procurement quantity 280 units; supply period 13 July 2025 to 14 July 2025." The draft sentence to be assessed is, "The procurement quantity in procurement plan P-0712 matches the group-order quantity in group-order result A-0712, and the supply period covers the day after cut-off." The retrieval request is "A-0712 leafy vegetable assortment procurement quantity group-order quantity supply period." The authorization boundary is read access to the public test copies of these three materials only, excluding signature pages and approval records.
Configuration and implementation checks
| Configuration item | Value or check action for this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | Associate only knowledge bases containing the three material copies and carrying the same vector model identifier | Check identifiers first and exclude mismatched bases |
| Search mode | Compare semantic, full-text, and hybrid retrieval separately against the test question | Identifiers A-0712 and P-0712 suit full-text or hybrid |
| Citation limit | For this hypothetical exercise, start at a token amount that can cover the key fields of all three materials, and compare against a smaller value covering only one material | Observe whether truncation drops the supply period |
| Minimum relevance | For this hypothetical exercise, start at a middle value and compare against a lower and a higher value | Takes effect only under semantic retrieval or reranking |
| Question optimization | Compare the original question with the optimized question and check whether subject and period were altered | A short question may introduce an unrelated subject |
| Result reranking | Confirm the rerank model is configured and enabled in settings | No rerank score is produced when it is not enabled |
Three common errors
- Associating knowledge bases that use different vector models. Check the base identifier and the vector model identifier, exclude mismatched bases, and record what was observed. Do not run a cross-model retrieval to compare missed recalls.
- Treating the minimum relevance value as a permission or entity filter. Use it only for relevance filtering under semantic retrieval or reranking. Check permission and entity isolation separately.
- Treating a draft sentence under review as an effective conclusion. Check whether the source text states signature, effective date, or authorization fields. When those fields are absent, record the sentence as unsupported or pending verification.
How to verify the setup
- Input the three source sentences from the hypothetical sample and the retrieval request. Compare whether the retrieval result contains A-0712, P-0712, and the supply period field together. If only a single material is returned, record incomplete recall; do not declare the source text contradictory.
- Input the draft sentence "The procurement quantity in procurement plan P-0712 matches the group-order quantity in group-order result A-0712." Compare the procurement quantity and group-order quantity values in the two source texts. If both are 280 units, the field agrees; if the values differ, record a same-field contradiction.
- Input the draft phrase "the supply period covers the day after cut-off." Compare the supply period stated in the procurement plan, 13 July 2025 to 14 July 2025, against the cut-off date of 12 July 2025. If the period start is later than the cut-off date, the phrase is supported; if the source text does not state a supply period, record it as unsupported. Separately, use a paper input of an empty array [] to check the downstream design for receiving empty citations. Do not claim that a fictional rule retrieval necessarily returns an empty array.