What the documents look like
Marketplace promotion registration and price management centers on four document types: activity registration forms, promotional price approval forms, inventory lock tables, and activity review tables. The core data is price, inventory, and campaign rhythm. The brief gives a weekly and per-campaign update frequency, and the daily work covers registration, approval, locking, and review. Product information and selection Q&A requires retrieving checkable fields from these materials, such as the registration status of a campaign, the signature status of an approval, the locked inventory quantity, and the price validity period. A retrieval result only provides quoted content. It does not constitute a submitted registration, an approved approval, or a deducted inventory quantity.
A practical difference from product title compliance work is that the center of gravity here is price, inventory, and campaign rhythm rather than promotional wording review. The typical practitioners are platform activity operators, brand sales planners, and finance business partners. Errors lead to campaign losses, removed resource slots, or price complaints. These facts shape what a retrieval step must expose: the original wording of a status field, the original wording of a signature field, and the original wording of a quantity field, so that a person can compare a proposed sentence against the source.
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 with different vector models cannot be associated together. Search modes include semantic retrieval, full-text retrieval, and hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and RRF is a fusion process rather than a toggle. The citation limit is measured in tokens. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used. Question optimization performs reference resolution and question expansion, and it adds an extra model call. Empty quoted content is output as an array whose length can be zero, and the downstream chain still runs.
A hypothetical paper exercise sample follows. The source text is: "The East China region A product activity registration form submitted on March 1, 2025 shows a registration status of pending review; the promotional price approval form has not been signed; the inventory lock table quantity is 0." The proposed sentence to be judged is: "The campaign registration has passed approval and the inventory has been locked." The source only shows pending review and unsigned. The proposed sentence claims approved and locked. These are mutually exclusive values on the same fields, so the proposed sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this exercise | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the same-vector-model base containing the registration form, approval form, lock table, and review table | Different vector models cannot be associated together; verify the vector model identifier first |
| Search mode | Test the same question separately with semantic retrieval and full-text retrieval | Compare the two modes against the test question; do not presume one mode suits all materials |
| Citation limit | Measured in tokens; test content completeness with materials of different lengths | Compare whether returned content is truncated, then adjust the limit |
| Minimum relevance | Label relevant and irrelevant samples first, then observe the filtering effect | This value takes effect only in semantic retrieval or with result reranking |
| Question optimization | Enable for follow-up questions containing references; disable for already explicit questions | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Use when a rerank model is available and enabled per settings | The rerank score can be used for filtering; it does not replace a permission judgment |
Three common errors
- Treating a retrieval hit as a completed approval or authorization. Input is the pending-review registration form above. Compare the registration status field with the proposed sentence. When retrieval returns the original text "pending review," the proposed sentence "has passed approval" is unsupported.
- Treating minimum relevance as an entity or permission filter. Input is an unrelated business question. Compare the returned content with the permission scope. A relevance score does not exclude unauthorized records; permission requires downstream implementation.
- Writing RRF as a toggle that can be switched on. Input is a hybrid retrieval configuration. Compare the documented description with the configuration item. RRF is a fusion process, and no corresponding toggle exists.
How to verify the setup
- Input the hypothetical source text and the proposed sentence above. Compare the registration status and approval signature fields. Observe whether the retrieved original text contains "pending review" and "not been signed." If it does, the proposed sentence is unsupported.
- Input an explicit paper empty result
[]. Compare whether the downstream design continues the chain with empty quoted content. Observe whether an error or interruption appears. - Input a follow-up question containing a reference. Compare the original question with the optimized question. Observe whether the subject was altered. If the subject changed, the question optimization result requires manual review.
The retrieval library same-vector-model requirement applies to knowledge base association. It does not extend to requiring the same model for business registration websites or other external systems. Semantic recall and relevance scores do not perform permission or entity isolation. Where a document field is not provided, the field remains unknown. A pending review does not prove the absence of an operation permission. Signing, cooperation reached, and contract effectiveness are three distinct facts. A quoted price validity period does not indicate a supply period. Do not invent authorization or effectiveness clauses to obtain a definite answer; record the evidence gap explicitly.