What the documents look like
Pinduoduo low-price campaign store operations center on four recurring documents: the campaign registration form, the product cost calculation sheet, the after-sales refund record, and platform internal messages. Each document carries its own field conventions. The campaign registration form states the campaign name, the registered product, the campaign price, and the campaign stock. The product cost calculation sheet states purchase price, packaging cost, express delivery cost, and gross margin calculation. The after-sales refund record states the order, the refund reason, and the refund amount. Platform internal messages state campaign rule notices and penalty notifications. The supplied brief gives an update frequency of daily and per campaign; this page does not derive batch synchronization or real-time mechanisms from that frequency.
Copy drafting draws its context from fields inside these documents. Campaign price, campaign stock, and refund reason belong to different fields of different documents. A drafted promotional sentence needs to indicate which document and which field it relies on. Token consumption comes from sending material fragments together with drafting instructions to the model. The reference limit is controlled in tokens, so longer individual fragments reduce how many fragments can be included at once.
A model's maximum context is the content length a single request can reference. The reply limit only constrains answer length and does not serve as the full context window. The reference limit uses tokens as its unit, and in a mixed knowledge base fragment lengths vary widely, so token-based control is more stable than item counts. A larger memory round count brings in more content, changing consumption and response time.
Constraints at the context limits step
The model's maximum context is determined by the actual model capability shown in the configuration dialog. The reply limit constrains only answer length. The reference limit is set in tokens, and whether material fragments are truncated needs to be checked against the total length of fragments plus drafting instructions. For single-turn drafting, a small amount of history may be retained; this is compared with the same question under empty history.
Hypothetical paper exercise sample: the material text is "Campaign registration form — May 2024 campaign: Product A, campaign price 9.9 yuan, campaign stock 500 units; Product cost calculation sheet — Product A: purchase price 6.5 yuan, packaging cost 0.3 yuan, express delivery cost 2.2 yuan; After-sales refund record — Order X: refund reason 'received item does not match description', refund amount 9.9 yuan." The sentence to be assessed is "Product A campaign price 9.9 yuan, sufficient stock, quality guaranteed." The assessment scope is the three materials above. The assessment action is limited to comparing the wording of the drafted sentence with material fields, and does not involve platform review or publication.
Configuration and implementation checks
| Configuration item | Value or check action on this page | Basis and comparison method |
|---|---|---|
| AI model | Select a model shown in the dialog as supporting the required capabilities | Model capability determines which configuration items are displayed |
| Maximum context | Check the value actually stated for the selected model | Compare with the total length of material fragments plus drafting instructions |
| Reply limit | A hypothetical trial starting value set to accommodate the basis and handling action | This starting value is a hypothetical trial value, not an official default; compare with the drafted sentence length |
| Reference limit | Set in tokens; check whether material fragments are truncated | Compare with individual fragment length and the question sample |
| Memory rounds | A small amount of history may be retained for single-turn drafting | Compare with the same question under empty history |
| Reply format | Keep default for ordinary copy drafting | Switch to json_schema only when a downstream process needs to read a fixed structure |
The maximum context value is checked against the model's stated capability. The reply limit starting value is chosen because the drafted sentence plus its basis needs to fit within the answer; it is a hypothetical trial value for this page and is compared against the drafted sentence length. The reference limit is compared against the total token length of the three material fragments plus the question. Memory rounds are compared by running the same question with and without one round of history.
Three common errors
- Treating the reply limit as the full context window. The reply limit constrains only answer length. When material fragments exceed the model's maximum context, they are still truncated, so the two values need separate checks.
- Using a relevance score in place of field comparison. The minimum relevance value is a filtering threshold and does not perform entity or permission isolation. Whether the drafted sentence is supported by materials depends on field-by-field comparison.
- Writing a pending-review status as already authorized. When the campaign registration form states the registration status is pending review, that status does not demonstrate campaign qualification and does not demonstrate the absence of operating permission.
How to verify the setup
- Input the three material texts and the drafted sentence from the hypothetical sample above. Compare whether "sufficient stock" in the drafted sentence and "campaign stock 500 units" in the campaign registration form share the same subject, the same period, and the same field. Compare whether "quality guaranteed" has a corresponding quality description field in any of the three materials. Only when a material states the corresponding field and the value matches is the wording supported. When no field is stated, record it as unsupported.
- Input the same drafted sentence once with empty history and once with one round of history. Compare whether the content actually sent to the model is identical in both runs, and observe whether history changed the referenced fragments. If the two inputs differ, record the differing fragments. Do not assert a specific failure cause from that difference.
- Input a paper-based empty-result downstream case: set the retrieval return to an empty list, check the downstream node's designed behavior under that input, and observe whether the downstream handles it as an empty reference. This setting is a paper assumption and does not indicate that actual retrieval will return an empty list.
The knowledge base reference limit uses tokens as its documented unit. Same-model retrieval association, empty-reference output, and context limits retain their documented conditions. A relevance score is not a category, entity, or permission filter. Retrieval is not authorization. The campaign registration form collects values; business verification, approval, database writes, and external synchronization require their stated downstream implementation. When the campaign registration form states a pending-review status, that status does not demonstrate campaign qualification and does not demonstrate the absence of operating permission. Missing evidence remains missing.