What the documents look like
Promotion enrollment and price management in this category runs on four recurring document types: the campaign enrollment form, the promotional price approval sheet, the inventory lock table, and the post-campaign review sheet. The operating focus is price, inventory, and campaign cadence. Review of promotional wording belongs to a different task and does not drive retrieval design here.
These materials are held by platform campaign operations, brand sales planning, and finance business partners. Updates arrive weekly and per campaign. In practice, one campaign produces several documents at once, with similar field labels and status words that are easy to confuse. A retrieval result reflects textual similarity between the query and stored content. It does not establish that a price was approved, that inventory was locked, or that a price has taken effect.
Constraints at the retrieval step
The knowledge base search node can be associated with one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be associated together. Confirm the vector model identifier first, exclude any knowledge base whose identifier does not match, and record what was observed. Do not run a cross-model retrieval in order to compare missed recalls.
Search modes are semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF. RRF is a fusion step in the pipeline, not a toggle that can be switched on. The reference cap is measured in tokens. The minimum relevance value applies only under semantic retrieval or when result reranking is used; it is a filter threshold, not a final ordering rule. Question optimization performs coreference resolution and question expansion, and it adds one model call before retrieval. The reference output is an array whose length can be zero, so the downstream path continues even when nothing is retrieved.
Hypothetical paper exercise. The source material reads: “Campaign enrollment form — March 2025 spring campaign, subject: brand sales planning, object: one platform store, status is pending review; promotional price approval sheet, same campaign, same field, status: approved; inventory lock table, same campaign, same field, status is pending lock.” The draft sentence under review is: “The campaign price is in effect, inventory is locked, and the resource slot enrollment can be submitted.” The source states only three status values: pending review, approved, and pending lock. It does not state a price-effectiveness clause or a completed inventory lock record. The draft sentence is unsupported and the evidence gap must be recorded.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only promotion enrollment knowledge bases sharing one vector model identifier | Confirm the vector model identifier matches before associating; exclude non-matching bases |
| Search mode | Test the same question first with semantic retrieval, then with full-text retrieval | Compare how campaign numbers and status fields are hit under each mode |
| Reference cap | Hypothetical starting value of 2000 tokens | Compare long and short enrollment forms and approval sheets to see whether content is truncated |
| Minimum relevance | Hypothetical starting value of 0.5 | Label relevant and irrelevant samples first, then compare filtered results |
| Question optimization | Enable only for multi-turn follow-ups and short questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled per settings | Compare candidate order before and after reranking and observe score changes |
Three common errors
- Treating a retrieval hit as an authorization. The input is an enrollment form with status pending review; the comparison object is the approval status field; the observable result is that retrieval returns the material. This does not mean approval is complete, and it does not mean the resource slot enrollment can be submitted.
- Associating knowledge bases that use different vector models. The input is two knowledge bases with different vector model identifiers; the comparison object is the model identifier; the observable result is a mismatch, so the base is excluded. Do not continue by describing a runtime outcome for the combined association.
- Treating minimum relevance as a permission or entity filter. The input is an unrelated business question; the comparison object is the similarity score; the observable result is that the score falls below the threshold and is filtered out. This does not mean unrelated businesses or unauthorized records were excluded.
How to verify the setup
- Input the draft sentence from the hypothetical exercise, “The campaign price is in effect, inventory is locked, and the resource slot enrollment can be submitted.” Compare it against the three status values in the the source material reads pending review, approved, and pending lock. The observable result is that price effectiveness and inventory lock completion are unsupported by the source text, so the sentence is marked as pending verification.
- Input the campaign enrollment form and the promotional price approval sheet from the hypothetical exercise. Compare the status value for the same field in the same campaign. The observable result is that the two status values differ, which is a mutually exclusive value pair for the same field, so it is a contradiction requiring manual review.
- Input an explicit paper-level empty result array. Compare it against the downstream handling design. The observable result is that the path continues to run, and the empty-result branch is recorded. Do not claim that a fictional rule retrieval necessarily returns an empty array.