What the documents look like
Marketplace promotion registration and price management in store service and internal training work around four record types: activity registration forms, promotional price approval forms, inventory lock forms, and activity review forms. The core operating facts are price, stock, and campaign rhythm. Store staff use these materials to answer questions about activity rules, price wording, stock status, and handover requirements.
For this category, the relevant difference from product title compliance work is that the review target is a price, a stock quantity, or a campaign stage. The four record types are updated weekly or per campaign. The same activity therefore appears at different stages: registration open, price pending approval, stock locked, or review completed. Retrieval results must let a reader tell those stages apart. A registration form states a deadline. A price approval form states a status. An inventory lock form states a locked quantity. A review form states outcomes after the campaign. None of these four record types substitutes for another.
Constraints at the retrieval step
The knowledge base search node can link one or more knowledge bases that use the same vector model. Knowledge bases built on different vector models cannot be linked together. The vector model identifier is checked first; bases whose identifiers do not match are excluded, and the actual observation is recorded. Retrieval is a recall process. It does not perform authorization decisions, and it does not execute registration, approval, or stock actions.
Semantic retrieval, full-text retrieval, and hybrid retrieval can be compared against test questions. No single mode is assumed to suit all four record types. Full-text retrieval suits identifiers, form numbers, and proper nouns. Semantic retrieval suits natural-language questions. Hybrid retrieval combines both through RRF, which is a result-fusion step inside hybrid retrieval and is not a separate switch.
The citation limit is measured in tokens. The minimum relevance value applies only under semantic retrieval or when result reranking is used. It is a filter threshold, not a final ordering rule. Question optimization performs coreference resolution and question expansion before retrieval; the original question and the optimized question are compared to check whether the subject was altered. Reranking requires a rerank model and is enabled according to settings. Image retrieval requires image input and matching model capability. Empty output is returned as an array whose length can be 0, and the downstream chain still runs.
A hypothetical paper exercise sample: the source text reads, "The East China store promotion registration form for March 2025 states that the registration deadline for Activity A is March 5. Promotional price approval form P-0312 is in pending-review status. The inventory lock form states that Activity A has 200 units locked." The draft sentence to be judged reads, "Activity A has completed promotional price approval and can proceed with registration." The subject under review is Activity A. The object is promotional price approval form P-0312. The status is pending review. Pending review does not constitute completed approval. The draft sentence does not match the source status, so it is unsupported, and the evidence gap is retained.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Linked knowledge bases | Registration forms and price approval forms only | Both bases share the same vector model identifier; bases with non-matching identifiers are excluded |
| Search mode | Hybrid retrieval | Compared with semantic retrieval and full-text retrieval on form numbers and natural-language questions |
| Citation limit | 1200 tokens | Hypothetical starting point for this exercise; compared with 800 and 1600 for content completeness |
| Minimum relevance | 0.35 | Hypothetical starting point for this exercise; relevant and irrelevant samples are labelled before adjustment |
| Question optimization | Enabled | Original and optimized questions are compared on multi-turn follow-ups to check the subject |
| Result reranking | Enabled | Ordering before and after reranking is compared when candidates are numerous; not enabled without a model |
Three common errors
- Treating a pending-review status as completed approval. The input is promotional price approval form P-0312 with pending-review status. The comparison object is the phrase "approval completed." The observable result is that the status field and the phrase do not match, so the draft is recorded as unsupported.
- Treating retrieval recall as authorization to execute. The input is a retrieved registration form fragment. The comparison object is a registration execution request. The observable result is that retrieval returns cited content only and produces no approval or registration action.
- Linking knowledge bases that use different vector models. The input is two bases with different vector model identifiers. The comparison object is the linking condition. The observable result is that the identifiers do not match, so the bases are excluded and the observation is recorded; no cross-model retrieval comparison is performed.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence "Activity A has completed promotional price approval and can proceed with registration." Compare the status field of the promotional price approval form. The condition is met when the observable result shows pending review and completed approval do not match, and the draft sentence is judged unsupported.
- Input a retrieval question that contains only the registration deadline for Activity A. Compare the returned cited content against the date field in the source text. The condition is met when the cited content contains March 5 and contains no statement of completed approval.
- Input a paper empty result array []. Compare whether the downstream node continues to execute. The condition is met when the chain runs through and produces no registration or approval action.
These checks describe expected observations. They do not claim that any test, deployment, approval, or business action has occurred. Where the source text does not state a field, the field remains unknown, and the gap is recorded rather than filled.