What the documents look like
Marketplace advertising adjusts bids in real time against traffic and conversion, so the review logic differs from platform promotion enrollment. The daily materials for this category include campaign plans, keyword reports, audience package descriptions, and conversion attribution tables. Data refresh occurs daily or hourly. A contract and transaction material comparison task has to check the subject, object, period, fields, and authorization state recorded across these documents. The retrieval step works on knowledge base content after the materials are ingested. The materials are field-dense and version turnover is fast. A knowledge base search node can associate one or more knowledge bases that use the same vector model. Search output is an array whose length can be 0. Retrieved results supply reference content only; they do not carry an authorization judgment.
Constraints at the retrieval step
Campaign materials update frequently. One campaign plan can exist in several period versions, and one keyword report can cover different statistical periods. Retrieval must first establish which version and which period was hit before field comparison begins. Full-text retrieval suits exact queries such as identifiers, model numbers, proper nouns, and error codes. Semantic retrieval suits natural language questions. Hybrid retrieval uses both recall paths and merges results through RRF. ReRank performs a second ordering of candidate text. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not handle permission or entity isolation. The reference limit is measured in tokens and controls the maximum content referenced per search. When question optimization is enabled by configuration, it performs coreference resolution and question expansion and adds one model call before retrieval. Knowledge bases with different vector models cannot be associated together; the model identifier must be checked and non-matching bases excluded. Cross-model retrieval is not executed to compare missed recalls. Retrieval is a read action after authorization, and neither the form nor the search node produces approval or signing effect.
Hypothetical paper exercise sample: the source material for checking reads, "Party A, a brand marketing department, and Party B, an agency operations service provider, signed an Advertising Placement Service Contract on March 1, 2025, agreeing that Party B would handle on-platform placement for an e-commerce platform from March to May 2025, and the contract takes effect from the date both parties affix their seals." A draft sentence to be judged reads, "The contract took effect on March 1, 2025, and Party B has obtained placement operation authorization." In this sample, the signing date, the effective condition, and the scope of authorization are three distinct facts. The draft sentence treats the signing date as the effective date and further derives operation authorization. Neither is supported by the source text.
Configuration and implementation checks
| Configuration item | Purpose on this page | How to check |
|---|---|---|
| Associated knowledge bases | Select bases holding campaign plans, keyword reports, audience package descriptions, and conversion attribution tables | Check whether the vector model identifiers of the bases match; exclude those that do not |
| Search mode | Compare semantic, full-text, and hybrid retrieval against test questions | Use identifier-type questions for full-text retrieval and natural language questions for semantic retrieval |
| Reference limit | Control the token volume referenced per search | Test with actual materials of different lengths for completeness, then adjust |
| Minimum relevance | Filter low-relevance results | Label relevant and irrelevant samples first, then tune the threshold and record the effective condition |
| Question optimization | Coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Second ordering of candidate text | Confirm the rerank model is available and enabled by configuration, and record the score distribution |
For the reference limit, measure the fragments to be referenced under the selected model first, then compare completeness under different limits; no length measurement exists at present. For minimum relevance, 0.5 may serve as a hypothetical comparison starting point, chosen for this exercise to compare against lower and higher thresholds on the labeled results; the reason is to have a mid-range value between permissive and strict filtering. Subject and permission are checked separately.
Three common errors
- Treating the signing date as the effective date. In the sample, the contract states it takes effect from the date both parties affix their seals, the signing date is March 1, 2025, and the seal date is not stated. The draft sentence says it took effect on March 1, 2025. The input is the signing date, the comparison object is the effective condition, and the observable result is that the effective date is unsupported. The correct action is to record the effective condition as pending verification.
- Treating a retrieval hit as authorization confirmation. The search node returns reference content and does not return a permission conclusion. The draft sentence says Party B has obtained placement operation authorization. The input is the retrieved contract text, the comparison object is the authorization clause, and the source states the two parties and the service period but provides no platform account permission or activation record. The correct action is to mark authorization as pending verification.
- Associating knowledge bases across vector models. If the vector model identifiers of two bases differ, they cannot be associated together. The input is the two model identifiers, the comparison object is whether the identifiers match, and the observable result is that a non-matching base is excluded. The correct action is to associate only bases with matching identifiers and to record the excluded items.
How to verify the setup
- Input the sample contract passage and the draft sentence, and check "takes effect from the date both parties affix their seals" together with the seal date. The source does not give the seal date, so the claim of taking effect on March 1 remains pending; if retrieval does not return that passage, record the reference gap first.
- Input the same source text and "has obtained placement operation authorization," and check the service duty and the platform account operation permission separately. The former has a record of handling placement from March to May; the latter has no corresponding permission or activation record, so operation authorization remains pending, and an empty retrieval result is not presumed.
- Use the sample title "Advertising Placement Service Contract" to compare full-text and semantic retrieval, and check item by item whether the returned passages belong to that contract. Record a miss when nothing is hit; no single mode is guaranteed to hit. For the empty-result path, use a paper input of [] to check the downstream design separately.