What the documents look like
Platform advertising adjusts bids against traffic and conversion in near real time, so the material used to judge a draft sentence is the same-day or same-hour keyword report and audience-package description. The working documents for this category are the delivery plan, the keyword report, the audience-package description, and the conversion attribution table. Their update rhythm is daily or hourly, which means a draft sentence can only be judged against the version of the material that was actually loaded into the conversation.
For multi-turn prompt work, the practical consequence is that the material scope, the answer task, and the handling of missing evidence must be written into the prompt. Because the source material changes often, a draft sentence that adds a price, a discount, an eligible audience, or a delivery window is not supported unless the loaded material states it. The model and length settings determine how much context and how long a reply a single turn can carry. The reply format is chosen according to what reads the output downstream, and when json_schema is selected, the corresponding format requirement must be filled in and the returned structure checked. Question optimization is only checked when a knowledge-base continuous conversation requires it, and then only for pronoun resolution and question expansion.
A hypothetical paper exercise sample is used throughout this page. The source material reads: "This delivery plan covers East China, the delivery window is daily 09:00 to 21:00, the featured product is a lightweight summer jacket, and the audience package is users who browsed similar products in the last 30 days." The draft sentence under review reads: "This product targets users nationwide, runs around the clock, and suits wear in all seasons." The subject is this delivery plan, the object is East China and the last-30-day similar-product browsers, the state is the delivery window and the featured product, and the effective scope is this plan. The words "nationwide," "around the clock," and "all seasons" do not appear in the material, so they are recorded as unsupported.
Constraints at the multi-turn prompts step
Multi-turn dialogue carries earlier turns forward, so the prompt must state the material scope, the answer task, and how to act when evidence is missing. Memory turns control how many earlier rounds the answer may reference; too few and the source material may no longer be present in the history, too many and more content is carried with added consumption and slower response. The reply limit caps how long a single answer can be. Temperature controls stability, and a lower value suits rule-bound judgment such as checking a draft sentence against source wording. Topp overlaps with temperature and is normally left alone when temperature already gives the wanted effect. Stop sequences end output at a fixed marker and are not needed for ordinary chat. The reply format stays at default for ordinary chat, and jsonschema is used only when a downstream step reads a fixed structure, in which case the format requirement must be supplied. Multimodal recognition depends on the model's own capability, and plain document files are not sent directly as multimodal input; their content still needs file parsing into text.
For this page's task, the prompt must not let the model supply business facts that the material does not state. The draft sentence above cannot be rewritten into a supported claim by inference; the missing scope, window, and audience remain missing.
Configuration and implementation checks
| Configuration item | Hypothetical starting value for this exercise | Reason and comparison method |
|---|---|---|
| Memory turns | Set to 4 turns for this exercise | The exercise needs the source material and the draft sentence to stay in view together; compare whether the history still contains the source wording at different turn counts |
| Reply limit | Set to 800 for this exercise | The draft sentence is short; compare whether the reply is cut off before the judgment is complete |
| Temperature | Set to 0.3 for this exercise | Draft judgment needs stability; compare whether the same draft sentence receives the same judgment at different temperatures |
| Reply format | json_schema selected for this exercise | A downstream step reads fixed fields; fill in the format requirement and check the returned structure |
| Multimodal recognition | Off for this exercise | The sample is plain text; compare whether enabling it introduces unrelated input |
These are exercise starting values, not official defaults. Each is compared against another value to see whether the observable result changes.
Three common errors
- Treating "nationwide" in the draft sentence as a delivery scope already stated in the material. The check inputs the source material and the draft sentence, compares "East China" and "nationwide" as the same field, and the observable result is that the draft value is unsupported.
- Setting memory turns so low that the source material no longer appears in the history. The check inputs multi-turn dialogue under one chatId, compares whether the history still contains the source wording at different turn counts, and the observable result is that the source wording is missing.
- Selecting json_schema without filling in the format requirement. The check inputs the draft sentence and the format requirement, compares the returned structure against that requirement, and the observable result is a structure mismatch.
How to verify the setup
- Input the source material and the draft sentence, compare the scope, delivery window, and audience in the draft against the same fields in the source, and the observable result is unsupported or pending verification. A value that is absent from the material stays absent; it is not a denial and not a conflict.
- Input multi-turn dialogue under one chatId, compare whether the history still contains the source wording at different memory-turn settings, and the observable result is that the source wording is missing at the lower setting.
- Input the draft sentence and the format requirement, compare the returned structure against the requirement, and the observable result is a structure mismatch. If the return is empty, use an explicit paper input of
[]to check the downstream design, without claiming that a fictional rule retrieval must return[].
A relevance score is not a category, entity, or permission filter, and retrieval is not authorization. Same-model retrieval association requires checking the identifier first and excluding libraries that do not match; it is not extended to external systems. A pending request is not approval, an offered option is not execution, and an unsigned termination is not a completed termination. Where the material does not state a value, the gap is recorded and left open.