What the documents look like
Promotion enrollment and price management centers on price, inventory and campaign cadence; advertising copy review belongs to a different workstream. The materials handled on this page are the campaign enrollment form, the promotional price approval form, the inventory lock form and the campaign review form. Sources include the platform marketing backend, the campaign backend, ERP inventory and historical transaction prices. The update frequency given in the brief is weekly and per campaign. Costing the lowest final price, locking campaign inventory and synchronizing prices across platforms are the time-consuming steps. Errors lead to campaign losses, loss of a resource slot, or price complaints. Document parsing and chunking must preserve the subject, object, period and status across the enrollment form, approval form and lock form so that contract and transaction material comparison can use them.
The brief describes a category, not an actual customer. Background marked as approximate stays conditional.
Constraints at the document chunking step
Chunking follows the structure of each material. Conditions and the explanations that qualify an amount, a quantity or a date stay in the same segment, so that a segment can still be traced back to its source text after splitting. In FastGPT, a knowledge base consists of library, collection and data. The smallest search unit is the library; a collection is only for classifying and managing data and does not affect search results. The dataset.datas table in MongoDB stores the vector source data, and an indexes field records the corresponding vector IDs, which is an array, so one set of data can correspond to multiple vectors. The vector field in PostgresSQL stores vectors. During retrieval, vectors are recalled first, then the source data is located in MongoDB by vector ID; when records correspond to the same set of source data they are merged, and the vector score takes the highest score. Semantic retrieval relies on vector distance; full-text retrieval suits numbers, models and proper nouns; hybrid retrieval is merged through RRF; the ReRank score falls between 0 and 1. The minimum relevance value takes effect only in semantic retrieval or when result reranking is used, and it is a filter threshold. The reference cap is measured in tokens, not top k. Retrieval does not perform authorization judgment. Forms collect values; business verification, approval, database writes and external synchronization each require their own downstream implementation.
Hypothetical paper exercise sample: source text one is "Brand A submitted on 1 March 2025 an enrollment for a single item campaign running 8 March to 10 March, enrollment form number B20250301, status pending review." Source text two is "Promotional price approval form P20250302 states the campaign final price for this item is 99 yuan, approval status pending review, valid until 10 March 2025." Source text three is "Inventory lock form S20250303 records 500 units locked for this item, lock period 8 March to 10 March." The draft sentence to be assessed is "The item campaign has been approved, the final price is 99 yuan, 500 units are locked, and listing can proceed." Assessment requires the same subject, the same period and the same field. Pending review does not constitute approval, and a quotation validity date does not indicate a supply period.
Configuration and implementation checks
| Configuration or check item | Hypothetical starting value and reason for this page | Comparison object and observation method |
|---|---|---|
| Search mode | Hybrid retrieval is used, because the materials contain numbers and natural-language conditions | Compare recalled content against semantic retrieval alone and full-text retrieval alone |
| Reference cap | Hypothetical starting value 3000 tokens, because a single enrollment form and approval form together are long | Compare with a smaller value and observe whether quoted content truncates a key condition |
| Minimum relevance | Hypothetical starting value 0.3, used to filter low-relevance fragments | Compare with a value such as 0.5 and observe whether a conditional sentence is filtered out |
| Question optimization | Enabled for multi-turn follow-ups, can be off for a single clear question | Compare the retrieval terms and recalled content before and after enabling |
| Chunk boundary | Split by material structure, with conditions and amounts in the same segment | Check whether a segment can be traced back to its source subject, period and status |
| Training and index state | After upload, check whether actual training and indexing have completed | Search separately with a term, a number and a natural-language question |
Three common errors
- Treating pending review as approved. In the hypothetical sample the enrollment form and the approval form both carry status pending review, while the draft sentence states the campaign has been approved. The input is the two pending-review materials, the comparison field is approval status, and the observable result is that the status values do not match.
- Treating the quotation validity date as a supply period. The approval form is valid until 10 March, and the draft sentence concludes from this that listing can proceed. The input is the validity field, the comparison object is the supply or listing authorization field, and the observable result is that this field is not stated.
- Treating a retrieval hit as authorized execution. The inventory lock form records 500 units and retrieval can recall that segment. The input is the locked quantity, the comparison object is the approval required for authorized listing, and the observable result is that authorization evidence is absent.
How to verify the setup
- Input the three source texts of the hypothetical sample and the draft sentence, compare the approval status field, and expect to observe that pending review and approved do not match, so the draft sentence is unsupported.
- Input the approval form validity date and the draft sentence listing conclusion, compare the validity field with the authorization field, and expect to observe that the authorization field is not stated, so the conclusion remains to be verified.
- Input the inventory lock form and the draft sentence that listing can proceed, compare the locked quantity with the approval authorization, and expect to observe that the lock record exists while authorization evidence is absent, so the gap must be recorded. An empty result can be checked against downstream design with the paper input
[]; this does not claim that any rule necessarily returns[]on retrieval. Libraries built on different vector models cannot be associated together; check the identifiers first and exclude those that do not match, then record the actual observation.