What the documents look like
Marketplace advertising adjusts bids in real time against traffic and conversion, and the specialist handling it works daily with campaign plans, keyword reports, audience package descriptions and conversion attribution tables. Copy drafting for products and marketing sits at the front of that chain: selling points, price wording and promotion scope written into copy are later used by the buyer for creatives and landing pages. Campaign data updates by day and by hour, so a promise written into copy that does not match the source material grows more expensive to correct as spend accumulates. This page addresses only the context and token step of drafting, meaning how much material the model can see, how much content is referenced, and how much prior conversation is carried in.
The documents behind this step are the campaign plan, the keyword report, the audience package description and the conversion attribution table. Copy drafting reads from them; it does not write back into them. Context and token decisions belong to the drafting interface, and the four documents supply the material that drafting must stay consistent with.
Constraints at the context limits step
Context and token involve three separate boundaries. The first is the length of content a single model request can reference, set by the selected model's maximum context. The reply cap limits output length only, and the two cannot substitute for each other. The second is the knowledge base reference limit, which controls in tokens how much content each search may reference; longer individual passages carry higher risk of being trimmed. The third is memory turns, which sets how many earlier conversation rounds an answer may reference; a larger number carries more content in and raises both consumption and response time.
Structured handoff needs separate checking. When the json_schema reply format is selected, downstream nodes depend on complete fields, and checking text length alone cannot reveal a missing field. Multimodal recognition depends on model capability; when the configuration dialog does not show the corresponding option, that capability is unavailable. Image search requires that an image index was built at ingestion time and that the search request carries the relevant image input and model capability; the paths shown in the overall flow diagram do not mean every request executes all of them. ReRanking requires a ReRank model enabled under its settings, and hybrid search merges multiple result paths through RRF. RRF is a fusion process, not a standalone switch.
Hypothetical paper exercise sample
The following is fictional for this paper check and points to no real customer or platform record.
Source text one: "This shop's spring skincare set, order during the promotion for an immediate 50 yuan reduction; the promotion period is subject to the shop announcement."
Source text two: "This set suits dry to normal skin; for oily skin, a patch test is recommended first."
Source text three: "Gift quantities are limited, while supplies last; gifts do not stack with other shop-wide spend-and-save offers."
Proposed draft sentence to be judged: "The spring skincare set suits all skin types, order during the promotion for an immediate 50 yuan reduction, and the gift is guaranteed to ship with the order."
Configuration and implementation checks
| Configuration or check item | Setting in this sample | How to check and what it means |
|---|---|---|
| AI model and maximum context | A model the dialog marks as handling longer material | The three source sentences are short; the main check is whether the model supports json_schema structured output |
| Reply cap | Set to a medium length as a hypothetical starting value for this exercise | This starting value was chosen because the draft sentence must express its basis and limits; compare against the combined length of the three source sentences to confirm no truncation |
| Knowledge base reference limit | Set in tokens to cover the three source sentences | Compare whether the referenced content fully includes the promotion scope and the skin-type limit; raise it if trimming appears |
| Memory turns | Set to a small number as a hypothetical starting value for this exercise | This starting value was chosen because the object being checked is a single draft sentence; compare against an empty-history input to confirm nothing extra is carried in |
| Reply format | Select json_schema when downstream reads fields | Check field completeness and successful parsing; checking text length alone is not sufficient |
| Minimum relevance | Takes effect only under semantic search or result reranking | This value is a filter threshold and does not carry permission or entity isolation |
Three common errors
- Treating the reply cap as the full context window. The input is the maximum context value shown in the model selection box together with the reply cap setting; compare whether the two are the same field. The observable result is that the two are recorded separately, and conflating them is a configuration-understanding error.
- Using a similarity setting to exclude unrelated material. The input is the minimum relevance value and the retrieval scope; compare whether that setting is described as filtering business ownership. The observable result is that relevance filters only low-scoring content and does not distinguish which entity a piece of material belongs to.
- Writing RRF as a switch that can be turned on. The input is the options listed in the search parameter interface; compare whether a standalone switch named RRF exists. The observable result is that RRF appears only as the fusion process of hybrid search, and the absence of a corresponding interface control marks the description as wrong.
How to verify the setup
- Verification one: for source text one and the phrase "order during the promotion for an immediate 50 yuan reduction" in the draft sentence, the input is both texts; compare whether the promotion-period wording matches. The observable result is that the source states the period is subject to the shop announcement while the draft sentence omits that limit, so the wording is unsupported and needs the limit added or a rewrite.
- Verification two: for source text two and the phrase "suits all skin types" in the draft sentence, the input is both texts; compare the value of the skin-type field. The observable result is that the source limits suitability to dry to normal skin and recommends a patch test for oily skin, while the draft sentence widens it to all skin types; the same field carries mutually exclusive values, which is a contradiction.
- Verification three: for source text three and the phrase "the gift is guaranteed to ship with the order" in the draft sentence, the input is both texts; compare the gift-status field. The observable result is that the source states limited quantities while supplies last, while the draft sentence changes this to a guarantee, so the wording is unsupported. If retrieval returns an empty result, a paper input of [] can be used to check how the downstream node is designed to handle an empty reference; no claim is made that any rule-based retrieval necessarily returns empty. Libraries built on different vector models cannot be associated together; check the identifier first and exclude those that do not match, record the actual observation, and do not infer missed recall or redundant results.