What the data for this category looks like
The core data for delivery effect optimization comes from advertising platforms, attribution systems, CRM leads, and landing page data, with updates on a daily or real-time basis. Daily processed documents include daily business reports, weekly review documents, ad adjustment records, and material A/B test data tables. Data fields include impressions, clicks, conversions, customer acquisition cost, budget balance, and others. Some files are semi-structured tables, others are plain text records, and there is no unified fixed format.
Constraints on Multi-turn Dialogue and Prompting Workflows
Real-time data updates require multi-turn dialogue to pull the latest data sources, avoiding reliance on stale content from static knowledge bases. Semi-structured document formats require prompts to clearly specify field parsing rules, preventing confusion between different types of business data. The multi-turn abnormal indicator positioning and budget allocation workflow requires retaining conversational history context to ensure subsequent follow-up questions can link to previous query content. High-frequency indicator query and adjustment requirements require prompts to streamline instructions and reduce invalid context usage.
How to Configure Settings
These configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
messages | Retain the last 20 rounds of conversation context, with no more than 1500 characters per round of content | Adapt to multi-turn follow-up scenarios for delivery optimization, avoiding context overflow |
variables | Reset variable cache with each request, only pass core indicators for the current session | Prevent variable result accumulation, aligning with real-time data update requirements |
| Maximum context length | 8000–12000 token | Accommodate complete content from multi-round adjustment records and review documents |
| Question optimization | Enable automatic optimization | Simplify the complexity of questions from ad optimizers, improving instruction accuracy |
chatId | Bind a unique session-specific identifier | Ensure session data from different optimizers is isolated from each other |
Three Common Misconfigurations
- The
variablesresult returned when calling the API accumulates historical content. This occurs because the variable context cache is not reset with each request, causing session data to accumulate continuously. - After enabling input guidance and thesaurus configuration, expected questions do not display in the conversation interface. This happens because the thesaurus is not bound to the current application's session scenario, or the display switch for input guidance is not enabled.
- After uploading an xlsx file, its content cannot be restated. This occurs because structured document parsing rules are not configured, or table chunk processing is not enabled during upload, resulting in the file not being parsed correctly.
How to Verify Proper Configuration
- Initiate a query containing core business indicators, and check whether the returned results match real-time or daily business data.
- Initiate two progressive rounds of questions, check whether subsequent questions can associate with the context from the previous round, and confirm that the
messagescontext is correctly retained. - Upload a test xlsx file, initiate a request to restate the table content, and confirm that the fields and values in the table can be returned correctly.
- After enabling the input guidance function, check whether preset guidance question options are displayed in the conversation interface.
The values provided are common starting points and should be measured against the user's own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.