What the documents look like
An MCN manages long-term account compliance for several creators at once, and its work does not serve a single brand session. The knowledge base for this task therefore holds four recurring document types: creator signing contracts, training records, violation handling forms, and account schedules. These materials update daily and weekly. A retrieval request during order fulfillment and after-sales question handling often crosses more than one type, for example whether a creator has a signed contract for a given session, whether a violation case affects that session's schedule, and which contract governs after-sales responsibility.
The knowledge base search node can link one or more knowledge bases that use the same vector model. If the four document types use different vector models, one node cannot link them together. The practical step is to check the vector model identifier, exclude the non-matching base, and record what was observed. Cross-model retrieval is not performed to compare missed recalls.
Constraints at the retrieval step
Retrieval returns cited content. It does not decide authorization. Whether a creator may start a live session, or whether an after-sales request is approved, depends on downstream steps such as the contract ledger and approval records. Cited content is output as an array, and its length can be zero. Even when nothing is found, that output path still runs.
Semantic retrieval relies on vector distance and suits natural-language questions. Full-text retrieval relies on keyword hits and suits exact queries such as contract numbers, violation form numbers, and account names. Hybrid retrieval uses both recall paths and merges them through RRF. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used. It is a filter threshold, not a ranking rule, and it does not act as a permission filter for creators or business records. Question optimization performs coreference resolution and question expansion and adds one model call before formal retrieval.
Hypothetical drill sample: an MCN operations staff member is preparing an answer to an after-sales follow-up. The source text available for checking reads: "Creator A signed a Live Cooperation Contract with the agency, contract number HT-2024-0317, covering the May 10, 2024 special session." "Creator A received violation handling form WX-2024-0091 on April 28, 2024; the handling status is pending review." "The account schedule shows Creator A's May 10, 2024 19:00 special session, with schedule status confirmed." The draft sentence to be judged reads: "Creator A's May 10 session has passed violation review and can go live normally." This draft sentence must be compared item by item against the source text.
Configuration and implementation checks
The table below lists settings already described in the official material, or implementation checks, tied to this sample and this task. The values are starting points for this hypothetical trial, used to compare against other values. They are not official defaults.
| Configuration item | Hypothetical trial starting point | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Four bases: creator signing contracts, training records, violation handling forms, account schedules | First check whether the four bases share the same vector model identifier; if not, exclude the non-matching base and do not link them together |
| Search mode | Hybrid retrieval | Compare against semantic retrieval alone and full-text retrieval alone, observing hit differences for contract number HT-2024-0317 and for natural-language questions |
| Citation limit | Set to 3000 tokens for this trial | Compare contracts and schedule materials of different lengths to see whether content stays complete, then adjust |
| Minimum relevance | Set to 0.35 for this trial | Label relevant and irrelevant samples first, then compare filtering effects only under semantic retrieval or result reranking |
| Question optimization | Enabled | Compare the original question with the optimized question, checking whether subjects such as "Creator A" and "May 10 special session" are altered incorrectly |
| Result reranking | Enabled as needed | After confirming a rerank model is available, compare candidate ordering before and after reranking; rerank scores can be used for filtering |
Three common errors
- Treating a retrieval hit as an authorization conclusion. For the input "can Creator A's May 10 session go live," retrieval may return the contract and the schedule, but violation handling form WX-2024-0091 has the status pending review. Pending review does not equal review passed, and the retrieval result does not constitute live-session authorization. The check compares the form's status field against the wording of the draft sentence.
- Forcing knowledge bases with different vector models into one link. If the four bases have inconsistent vector model identifiers, the node cannot link them together. The correct step is to check the identifiers, exclude the non-matching base, and record the observation. Do not invent missed recalls or redundant results from a combined link.
- Treating minimum relevance as a permission or entity filter. This value filters low-relevance results and does not distinguish creator or agency ownership. For the input "Creator B's after-sales responsibility," if the knowledge base also holds Creator A's materials, similar content may still be recalled. The check confirms whether the subject of the cited content is Creator B.
How to verify the setup
- Enter the source text above together with the draft sentence, and compare the status field "pending review" on violation handling form WX-2024-0091 against the draft wording "has passed violation review." If the cited content shows that field as pending review, the draft sentence is not supported, and the item is recorded as pending verification rather than as passed.
- Enter contract number HT-2024-0317 and observe it separately under hybrid retrieval and semantic retrieval alone. If the full-text or hybrid path hits the contract containing that number while semantic retrieval alone does not, this shows that exact-number queries depend on the keyword path. Both results are recorded; a first-position hit is not guaranteed.
- Enter an after-sales request with no corresponding material, for example "Creator C's June 1 session refund request." Observe whether the cited content array is empty. If it is empty, the downstream design for handling an empty result can be checked. Do not claim that a fictional rule guarantees an empty array.
- Check the vector model identifiers of the four knowledge bases. If the identifiers match, they can be linked together. If they do not match, exclude the non-matching base and record the actual observation. Do not run cross-model retrieval to compare missed recalls.
- With question optimization enabled, compare the original question with the optimized question. If the optimized subject changes from "Creator A" to another creator, record the difference and check it. Do not assume the optimized result is correct by default.
- Compare citation completeness for contract and schedule materials when the citation limit is set to 3000 tokens and when it is set lower. If a lower value truncates key contract clauses, record that observation, then decide whether to adjust.