What the documents look like
Platform advertising adjusts bids against traffic and conversion in near real time, and campaign registration review is not the controlling gate. The working materials for this category are campaign plans, keyword reports, audience package descriptions and conversion attribution tables. These materials update daily or hourly, so two people in a handover can read different values from the same field at different moments. Citation tracing requires every stated conclusion to return to specific source text, and a missing material is recorded as a gap.
The knowledge base search node outputs citations as an array, and the array length may be zero. An empty array does not mean the workflow has failed and terminated. That property matters during handover, because a reviewer who receives an empty citation array must inspect the downstream design rather than assume an error.
Constraints at the citation tracing step
The search node can associate one or more knowledge bases that share the same vector model. Knowledge bases with different vector models cannot be associated together. During handover, the reviewer first checks the vector model identifier of each candidate base and excludes those that do not match.
The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filtering threshold and does not perform permission isolation or entity isolation. The citation cap is controlled by tokens, and it is a separate field from the thinking budget. Image retrieval requires image input and a model with the corresponding capability. Reranking requires a rerank model and must be enabled in configuration. The optional paths shown in the overall retrieval diagram are not executed on every request.
Hypothetical drill sample: the source text reads, "Party A signed the Campaign Handover Confirmation on 1 March 2024, confirming the transfer of Account A's keyword report and audience package description to Party B; the authorization scope is limited to routine bid adjustment and excludes budget increases. The confirmation states that Party B must complete the first review round by 15 March." A draft sentence under review reads, "Party B has been authorized to increase the budget and has completed the first review round." The draft and the source carry mutually exclusive values on the authorization scope field. The signing fact does not imply budget increase authority, and the review deadline does not mean the review is complete.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting point for this drill | Reason and comparison object |
|---|---|---|
| Associated knowledge bases | Only bases sharing one vector model | Compared with the vector model identifiers of other candidate bases |
| Search mode | Hybrid retrieval | Compared with recall range when semantic retrieval is used alone |
| Minimum relevance | Hypothetical starting value 0.5 | Compared with result count when filtering is not enabled |
| Citation cap | Hypothetical starting value 2000 tokens | Compared with citation count at a lower value |
| Citation output | Accept the array, length may be zero | Compared with whether the downstream step errors on an empty array |
| Material version check | Current version placed beside historical version | Compared with checking sources against the historical version alone |
The minimum relevance value, the citation cap and the thinking budget are distinct fields. The relevance score is not a category filter, an entity filter or a permission filter, and it does not exclude unrelated businesses or unauthorized records. Retrieval is not authorization.
Three common errors
- Treating a retrieval hit as executed authorization. Input: the hypothetical source text above and the draft sentence. Comparison: the authorization scope field. The draft writes a budget increase; the source limits scope to routine bid adjustment. The two values are mutually exclusive on the same field, so the draft is unsupported.
- Treating a pending review as missing operating permission. Input: the confirmation's statement that the first review round must be completed by 15 March. Comparison: the review status field. The source records only a deadline and records no completion, so the claim of completion is unsupported.
- Associating knowledge bases with different vector models. Input: the vector model identifiers of two bases. Comparison: whether the identifiers match. When they do not match, the base is excluded. A cross-model retrieval is not executed to compare missed recall.
How to verify the setup
- Input the hypothetical source text and the draft sentence. Compare the authorization scope field values. When the draft value is mutually exclusive with the source value, record the draft as unsupported and do not supply an authorization clause.
- Input the hypothetical source text. Compare the review status field. When the source carries only a deadline and no completion record, record an evidence gap and do not treat the review as complete.
- Input the vector model identifiers of two knowledge bases. Compare identifier consistency. When they differ, exclude the non-matching base and record the actual observation. Do not assert what would happen if the two were associated.
For the citation output, a paper input of an empty array can be used to check the downstream design without claiming that a constructed rule returns an empty result. Each verification question compares one field across the same subject and the same period. A missing record stays missing, and a pending request is not an approval.