What the documents look like
MCN management covers multiple creators under long-term account compliance, so the material set differs from a single brand campaign. The documents involved in operating reviews and exception review are creator signing contracts, training records, violation handling forms, and account schedules. These materials update daily and weekly. A review needs to retrieve the signing status, training completion, violation handling progress, and schedule arrangement of a specific creator from that material, so that reported operating figures and exception causes can be checked against the original records. Retrieval results feed directly into review conclusions, and consequences such as account bans, cooperation termination, or institutional compensation make accurate association with the correct creator record important.
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; the vector model identifier must be checked first, non-matching bases excluded, and the actual observation recorded. Semantic retrieval, full-text retrieval, and hybrid retrieval are compared against test questions; no single mode is presumed suitable for all material. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. The reference limit is measured in tokens. Question optimization performs coreference resolution and question expansion; the original question and the optimized question are compared to check whether the subject was altered incorrectly. The output reference content is returned as an array and its length can be 0, meaning the output path still completes even when no content is retrieved.
Hypothetical drill sample: an MCN knowledge base holds a creator signing contract with the text "Creator A signed an exclusive management contract with the MCN agency on March 1, 2025, with a term from March 1, 2025 to February 28, 2027. Creator A's account schedule is three live streams per week, arranged on Monday, Wednesday, and Friday." A violation handling form holds the text "Creator A used violating wording in a live stream on June 10, 2025; handling status is pending review." A draft sentence to be assessed reads "Creator A's contract was terminated on June 10, 2025." The draft sentence is checked against the term field in the contract text.
Configuration and implementation checks
| Configuration item | Hypothetical trial starting value | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Creator contract base, violation handling form base | Both bases use the same vector model; the model identifier is checked before associating them together |
| Search mode | Hybrid retrieval | Contract terms and violation status contain dates and status words; hybrid retrieval uses both semantic recall and full-text recall |
| Reference limit | 800 tokens | Contract clauses and violation descriptions are of medium length; compared with 400 tokens and 1600 tokens for content completeness |
| Minimum relevance | 0.5 | Takes effect only in semantic retrieval or result reranking; compared with 0.3 and 0.7 for filtering of relevant and irrelevant samples |
| Question optimization | Enabled | Review follow-up questions often contain references; the original question and optimized question are compared for subject consistency |
| Result reranking | Enabled | When candidate results are numerous, text is reordered; ordering before and after reranking is compared |
Three common errors
- Treating a violation handling form under pending review as completed handling. The form text states "handling status is pending review." Pending review does not demonstrate a lack of operating authority, and it does not mean handling is complete. The check compares the status field in the handling form with the status description in the draft sentence.
- Treating the contract signing date and the contract termination date as the same field. The contract text states a term from March 1, 2025 to February 28, 2027, while the draft sentence states "terminated on June 10, 2025." These are different fields. The check compares the termination date in the draft sentence with the term end date in the contract text.
- Associating knowledge bases that use different vector models. Before association, the vector model identifier of each base is checked, non-matching bases are excluded, and the actual observation is recorded. Cross-model retrieval is not executed to compare missed recall.
How to verify the setup
- Input the contract text from the hypothetical drill sample and the draft sentence "Creator A's contract was terminated on June 10, 2025," then compare the termination date in the draft sentence with the term end date of February 28, 2027 in the contract text. If the two dates differ, the draft sentence conflicts with the contract text, and either the draft sentence is corrected or contract change evidence is added.
- Input the violation handling form text from the hypothetical drill sample and a pending request "Creator A's violation has been fully handled and scheduling can resume," then compare the status field "pending review" in the handling form with "fully handled" in the request. If the status field is pending review, the completion status in the request is unsupported, and a decision waits until review is complete.
- Input the contract text from the hypothetical drill sample and the question "Creator A's schedule arrangement," then compare whether the retrieved reference content contains the schedule field "three live streams per week, arranged on Monday, Wednesday, and Friday" from the contract text. If the reference content contains that field and the subject is Creator A, the retrieval hit is recorded; if the reference content is an empty array, the miss is recorded and whether the data has completed training is checked.