What the documents look like
Group-buy presale operations collect orders first and organize procurement and fulfillment afterward. The working materials for operating reports and exception reviews are the presale product list, the order consolidation result, the procurement plan, and the cut-off notice. These documents update daily or per sales session. The cut-off notice and the consolidation result meet at a point in time, and the procurement plan depends on the consolidated order quantity produced after cut-off. Retrieval at this step must bring materials for the same session, the same product, and the same cut-off moment together so a person can compare them.
A retrieval hit is not an authorization to act. A draft procurement plan, a pending review state, and a sent cut-off notice are separate facts. The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases with different vector models cannot be associated together. The minimum relevance value is a filter threshold, and it does not isolate sessions, products, or permissions. The citation limit is measured in tokens.
Constraints at the retrieval step
Search runs in semantic, full-text, or hybrid mode. Hybrid search combines multiple result paths through RRF, which is a fusion process rather than a switch. The minimum relevance value takes effect only in semantic search or when result reranking is used. Reranking requires a rerank model and must be enabled according to settings. Image retrieval requires image input and matching model capability. These optional paths do not run on every request.
When no content is found, the citation output is an array whose length can be zero, and the downstream output path still runs. Cross-model association is not performed; check the model identifier of each base first and exclude those that do not match. Do not run cross-model retrieval to compare missed recalls.
Hypothetical drill sample: the source material states, "Presale session A closes at 18:00 on March 10, and the cut-off notice has been sent. The consolidation result records that a certain product in this session reached a group quantity of 120 units. The procurement plan draft states that procurement for this product is 100 units." The draft sentence to be judged is, "Procurement for this product in this session has been placed at 120 units." The fields needed for judgment are the actual procurement quantity and the procurement placement status in the procurement plan.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the base holding this session's operating documents | Check each base's vector model identifier first, exclude non-matching ones, and do not run cross-model retrieval |
| Search mode | Hybrid search | Run the same test question under semantic, full-text, and hybrid modes, then compare recalled content |
| Citation limit | Set to a token amount that covers one complete cut-off notice | Test with cut-off notices and consolidation results of different lengths to see whether content is truncated |
| Minimum relevance | Label relevant and irrelevant samples first, then set the value | This value takes effect only in semantic search or with reranking; adjust against the labeled results |
| Question optimization | Enable for follow-up questions containing references | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enable when candidates are numerous | A rerank model must exist and be enabled according to settings; use rerank scores to assist filtering |
Three common errors
- Treating a retrieval hit as authorized execution. Input: the procurement plan draft stating 100 units and the draft sentence claiming placement at 120 units. Compare the procurement quantity field and the procurement placement status field. The observed result is a quantity mismatch and an unstated placement status, so the draft sentence is unsupported.
- Treating minimum relevance as an entity or permission filter. Input: the same test question and documents from different sessions. Compare whether recalled content mixes in other sessions. The observed result is that the relevance setting does not distinguish session ownership, so document identifiers must be checked separately.
- Treating a pending review state as missing operating permission. Input: the cut-off notice and the procurement plan. Compare the notice send time with the plan review state. The observed result is that the notice has been sent while the plan remains a draft; these are different facts and cannot be inferred from each other.
How to verify the setup
- Input the sample source material and the draft sentence, then compare the procurement quantity and procurement placement status fields. The draft sentence is supported only when both fields are stated and consistent; in this sample it is unsupported.
- Input the same test question under semantic search and full-text search, then compare the two recall paths. Observe whether hybrid search, after RRF fusion, produces richer candidates. A first-position hit or full recall is not guaranteed.
- Input an explicit paper-form empty result array and check how downstream handles an empty citation. Do not claim that a fictional rule retrieval necessarily returns an empty array.