What the documents look like
Promotion enrollment and price management turns on price, inventory and campaign rhythm rather than on advertising wording. The working materials for this category are the campaign enrollment form, the promotional price approval sheet, the inventory lock sheet and the campaign review sheet. They draw on the platform marketing console, the campaign console, ERP inventory and historical transaction prices. In a cross-border setting the same campaign produces Chinese and English materials at the same time, and one price field may be worded differently across language versions. Parsing and chunking quality decides whether retrieval can restore a specific campaign, a specific price field and a specific validity period accurately. The materials are also the ones that carry authorization state: an approval sheet records whether a price is pending review or approved, and a lock sheet records whether inventory has been reserved. Retrieval and forms collect or return values; they do not grant approval, release inventory or execute a campaign.
Material identities matter at this step. A campaign name, a price field, a validity period and an approval status belong to one document version. If chunking separates a price condition from the scope it constrains, retrieval can return a fragment that looks complete while the governing condition sits in another chunk. Original data and retrieval indexes are separate layers: one group of original data can map to several indexes, retrieval hits an index, and pointing back to the source still needs document identity and version. Semantic retrieval supports cross-language understanding and suits natural-language questions; identifiers, model numbers and proper nouns are better served by full-text or hybrid retrieval. Hybrid retrieval merges multiple result paths through RRF, and reranking participates only when a model is available and the setting is enabled. Similarity and minimum relevance are filtering and ordering parameters; they do not perform permission or entity isolation.
Constraints at the document chunking step
Chunking has to follow the structure of each material. Conditions, notes and the content they constrain stay in the same segment, so a price condition in an approval sheet is not split from its scope of application. A campaign enrollment form and a price approval sheet have different field layouts, and an English enrollment form may order fields differently from its Chinese counterpart. The chunking step has to preserve enough of that structure for a later comparison to be possible.
The second constraint is that a retrieval hit is not a business fact. A returned chunk can show a price field and a validity period while the approval status in the source remains pending review. A pending request is not an approval, an offered option is not execution, and an unsigned termination is not a completed termination. The chunking step cannot resolve that gap; it can only avoid making it worse by keeping status fields attached to the content they qualify.
The third constraint concerns updates. Materials are refreshed weekly and per campaign, and after a change the actual training and index state has to be checked. A successful upload is not evidence that retrieval has been updated. Whether a new version is being returned is an observation to make, not something the upload action establishes.
A hypothetical paper exercise sample is used below. The source text reads: "Campaign name: Autumn full reduction; enrollment entity: a flagship store in East China; promotional price field: net price 199 yuan; validity period: 2025-09-01 to 2025-09-30; approval status is pending review." The draft sentence to be judged reads: "The store's Autumn full reduction net price of 199 yuan has been approved and can be executed in September." The source states only that the approval status is pending review. It does not state that approval was granted, and it does not state that execution is permitted. The draft sentence is not supported.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason for the choice and comparison method |
|---|---|---|
| Search mode | Hybrid retrieval | The materials contain Chinese natural language and English identifiers, so both semantic and keyword hits need coverage; compare against semantic-only and full-text-only on the same sample |
| Reference limit | 2000 tokens as a hypothetical starting point for this exercise | Enrollment forms and approval sheets carry many fields; compare against a lower starting point to observe completeness of the returned reference content |
| Minimum relevance | 0.3 as a hypothetical starting point for this exercise | Compare against a higher starting point to observe whether valid price fields are filtered out |
| Question optimization | Enabled | Multi-turn follow-ups often omit the campaign name; compare recalled content with the setting on and off |
| Result reranking | Enabled when a model is available | When candidates are numerous, compare ordering before and after reranking; the rerank score is 0-1 |
| Document identity and version | Checked against the material | Compare the source structure with the parsed result and retain an identity and version that can point back to the source |
Three common errors
- Treating the same price field worded differently across language versions as one fact. Check: input the Chinese approval sheet and the English enrollment form, compare the entity, campaign name, price field and validity period, and observe whether the same field shows mutually exclusive values. Parts not stated are recorded as pending verification.
- Treating a successful upload as an updated retrieval index. Check: after a material change, input the original campaign name and identifier, compare whether the returned content points to the new version, and observe whether the old segment is still hit. Training and index state are recorded from actual observation.
- Treating a retrieval hit as completed authorization or approval. Check: input the draft sentence from the hypothetical sample above, compare the retrieval result with the approval status in the source, and observe whether the wording "has been approved" appears. Retrieval does not produce an approval conclusion.
How to verify the setup
- Input the hypothetical sample source text and the draft sentence, compare "has been approved" in the draft with "pending review" in the source, and observe whether the retrieval result retains the pending-review status and does not output an approved conclusion. The condition is met when the status is consistent.
- Input the campaign name "Autumn full reduction" and the price field "net price 199 yuan", compare whether the returned segment contains the entity, the field and the validity period, and observe whether the condition and the content it constrains stay in the same segment. The condition is met when the field is complete.
- Input the explicit paper empty-result input
[], compare how the downstream design handles an empty reference, and observe whether it takes the empty-result branch. Do not claim that a fabricated rule makes retrieval return[]in every case.
These checks describe what to compare and what remains unknown. They do not assert that a test, deployment, approval, refund or other business action has occurred.