What the documents look like
For product questions in MCN operations, this page examines FastGPT retrieval. Compare the material received with the sources used in the answer. The examples are hypothetical.
Constraints at the retrieval step
The product-selection question starts with the category scope listed in the proposed influencer contract. Test a literal category term and a natural-language question against that same source, comparing whether each result retains the whole scope. Semantic, full-text and hybrid searches need a common document baseline. A matching product word alone is insufficient to explain its surrounding restrictions. If account schedules are also searched, check the associated knowledge bases' embedding models and inspect whether scheduling text displaces the contractual evidence.
Configuration and implementation checks
| Check | Task-specific condition | Evidence to retain |
|---|---|---|
| Associated knowledge bases | Proposed contract and account schedule; matching embedding models across bases | Actual source scope |
| Search mode | Compare semantic, full-text and hybrid using the same contract | Category clause retained with the beauty-product wording |
| Citation limit | Token-based; select using a long-clause test | Both the restriction and digital-stream schedule survive |
| Minimum relevance | Range 0–1; active for semantic search or reranking | Labeled category evidence and unrelated commission clause |
| Question optimization | Compare complete original and optimized questions | Category scope has not become model-specific permission |
Documented conditions
| Documented item / implementation check | Condition |
|---|---|
Associated knowledge bases | Knowledge bases associated with one search node must use the same embedding model. |
Search mode | Semantic, full-text and hybrid modes are available. |
Citation limit | Citation limits are measured in tokens. |
Minimum relevance | Range 0–1; active for semantic search or reranking. |
Question optimization | Resolves references and expands questions. |
Three common errors
- If a beauty-only contract becomes a product noun in the citation, compare the full clause with actual citation content to locate the missing restriction.
- If a digital-product schedule becomes contractual permission, preserve both sources. A scheduled category does not establish an amended contract.
- If a commission clause supports a claimed model approval, inspect candidate relevance to that model question and whether optimization added an unsupported product entity.
How to verify the setup
A separate test contract supplies a commission rate without model-specific promotion requirements. A returned commission clause still needs comparison with the question; related text cannot establish permission for a particular model. Retain the missing requirement, retrieved passages and claims made in the answer. After obtaining the relevant source, repeat the question and inspect whether question optimization preserves the product subject. Record the document addition separately from search-setting changes so that the change in the answer's evidence can be explained.
| Material context | Scope |
|---|---|
| Document types | Influencer Contract, Training Records, Violation Handling Form and Account Scheduling |
| Candidate sources | MCN Backend, Platform Violation Notice, Influencer Account Data and Contract Ledger |
| Technical baseline | FastGPT v4.17.2 |