What the documents look like
An MCN organization manages long-term account compliance for several creators at once. The working materials are creator signing contracts, training records, violation handling forms, and account scheduling sheets. For product and marketing copy drafting, the retrieval step draws citable wording from these materials: the cooperation scope agreed in a contract, product descriptions recorded in training, and prohibited expressions listed on a violation handling form. Materials update daily or weekly. Retrieval runs per request input; the update frequency does not trigger retrieval by itself.
The relevant category difference is that MCN management covers many creators across ongoing accounts, so a retrieved passage carries a creator identity, a status, and a validity window, not just a brand session. A retrieved sentence about one creator does not apply to another.
Constraints at the retrieval step
One retrieval combines semantic search, full-text search, hybrid search, question optimization, and reranking according to the input content and search parameters, then fuses the paths into reference content. The reference content is output as an array whose length can be 0, so the output path still runs when nothing is found. Associated knowledge bases must use the same vector model; bases with different vector models cannot be associated together. During review, check the model identifier of each base, exclude those that do not match, and record the actual observation. Do not run a cross-model retrieval to compare missed recalls.
Hybrid search merges vector and full-text results through RRF, and reranking applies a second ordering when a rerank model is available and enabled by configuration. Image retrieval requires an image input and the corresponding model capability; it is not part of every request. Reranking likewise is not part of every request.
A relevance score is a similarity value. It does not act as a permission filter, an entity filter, or a business-scope filter, and it does not exclude unauthorized records. Retrieval is not authorization.
Hypothetical paper exercise sample. The material text is: "Creator A signing contract, clause 3: during the cooperation period, product photos provided by the brand may be used; medical efficacy statements may not be used." "Training record 2024-05-10: Creator A has studied the product ingredient description; efficacy wording has not been confirmed." "Violation handling form 2024-06-02: Creator A received a platform warning for using the term 'treatment' in a livestream; handling status is pending review." The draft sentence to be judged is: "Creator A may claim in a livestream that the product has a therapeutic effect." The subject under review is Creator A, the object is product efficacy wording, the status is violation handling pending review, and the validity scope is the cooperation period. This sample is constructed for this page and was not tested.
Configuration and implementation checks
| Configuration item | Setting or check relevant to this page | Relation to the sample |
|---|---|---|
| Associated knowledge bases | One or more bases using the same vector model may be associated | Check each base's vector model identifier first and exclude mismatches |
| Search mode | Compare semantic, full-text, and hybrid search on the test question | The sample contains clause numbers and proper terms, so full-text and hybrid can be compared |
| Reference limit | Unit is tokens; test completeness against actual material length | Hypothetical starting value set to cover the three material passages; compare with a value shorter than that |
| Minimum relevance | Takes effect only in semantic search or when reranking is used | Label relevant and irrelevant samples first, then adjust the threshold |
| Question optimization | Used for coreference resolution and question expansion | Compare the original question with the optimized question and check whether the subject was altered |
| Reranking | Applied to candidate ordering when a model is available and enabled by configuration | When candidates are numerous, compare order before and after reranking |
The hypothetical reference-limit starting value is based on the combined length of the three material passages. It is a starting point for this paper exercise, not an official default value.
Three common errors
- Treating a semantic recall or a relevance score as a permission or entity isolation mechanism. The erroneous input is returning Creator B's contract content on the strength of similarity alone. The comparison is between the subject identifier on the returned record and the subject of the retrieval request. The observable result is that the returned content's subject differs from the request subject.
- Treating minimum relevance as a sorting rule. The erroneous input is the assumption that raising the threshold places the most relevant entry first. The comparison is between the entry set and order before and after filtering. The observable result is that filtering only reduces entries; order is still determined by fusion and reranking.
- Associating bases that use different vector models. The erroneous input is selecting bases with differing vector models together. The comparison is between each base's vector model identifier. The observable result is that bases with mismatched identifiers are excluded and do not enter this retrieval.
How to verify the setup
- Input the three material passages and the draft sentence from the hypothetical sample. Compare the efficacy wording in the draft sentence against clause 3 of the contract, the training record, and the violation handling form. The observable result is that when the draft sentence conflicts with the material record, it is not supported by the reference content, and the violation handling status remains pending review.
- Input a retrieval question containing an explicit number. Compare the entries returned by full-text search and by hybrid search. The observable result is that an entry where the number is matched appears in the reference content; when no match occurs, the reference content is an empty array.
- Input an empty reference content array. Compare how the downstream design handles the empty array. The observable result is that the downstream processes the empty array as such, with no fabricated entries. This check does not assert that any rule-based retrieval necessarily returns an empty array.