What this category's data looks like
Typical users of this category are ad optimizers, growth analysts, and agency operations teams. Daily processed documents include daily reports, weekly reports, price adjustment records, creative test sheets, and optimization retrospectives. Data sources cover ad platforms, attribution systems, CRM leads, and landing page data. Update frequencies are mostly daily or real-time. Data fields include ad delivery cost, conversion count, click-through rate, budget balance, and more. Units include yuan, count, percentage, and more. The content length of individual documents varies widely: some are short price adjustment records, while others are full-cycle delivery retrospectives spanning thousands of characters.
Constraints for Context and Token Management
The need to associate multi-source data requires context to cover historical records from different systems, raising requirements for recall accuracy. Real-time updated ad data means context must have timeliness. Excessively long context will cause token usage to exceed limits, leading to model call timeouts. The diversity of document structures—both structured reports and unstructured retrospective notes—requires context extraction to adapt to content in different formats. The most time-consuming step of abnormal indicator locating requires rapid recall of relevant historical data. If the context recall range is too large, it will increase token consumption and processing delay, and may introduce irrelevant data causing context pollution. Unlike ad delivery plan setup work, ad delivery effect optimization context needs to associate full-cycle operational data after ad launch, further increasing the complexity of context management.
How to set configurations
Based on the parameter logic of FastGPT v4.17.0, the following configurations are recommended:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
maxContext | First 20-30 entries | Matches the number of historical operation records that ad optimizers typically review, avoiding redundant data occupying excessive tokens |
quoteMaxToken | 8000-12000 token | Adapts to the content length of long documents such as ad daily reports and creative test sheets, ensuring key business fields are not truncated |
| Context recall threshold | 0.75-0.85 | Filters low-relevance historical data, reducing the probability of context pollution |
| Real-time data refresh interval | 5 minutes | Matches the real-time data update rhythm of ad platforms, ensuring context timeliness |
| Maximum length of single context entry | 1000 characters | Adapts to short-field content such as price adjustment records and conversion details, avoiding single entries occupying excessive token quotas |
Three common misconfigurations
- Incorrect number of context entries displayed: Set
maxContextto 30 entries, but only 2 entries appear in the conversation details panel. The cause is failure to correctly associate historical conversation nodes from ad platforms, only recalling content within the current conversation without covering cross-session historical operation records. - Abnormal JSON format in AI responses: AI-generated replies include non-standard JSON structures. The cause is that
quoteMaxTokenis set too small, resulting in key structured data fields being truncated and breaking the complete JSON structure. - Long task execution timeout: Submitting an optimization retrospective task with multiple long documents triggers an execution timeout error. The cause is failure to limit the token length of single context entries, causing total context tokens to exceed the model's call limit.
How to confirm correct configuration
- View the conversation details panel, confirm that the number of recalled context entries matches the
maxContextsetting value, and covers cross-session historical operation records. - Import a typical ad daily report document, trigger an AI-generated reply, and verify that the reply contains complete structured data with no field truncation.
- Submit a test task containing multiple long documents, confirm that execution does not produce timeout errors, and token usage does not exceed the model's publicly listed limits.
- Simulate an abnormal indicator locating scenario, enter relevant keywords, and check that the relevance of recalled historical data meets the preset threshold requirements.
The values provided are common starting points and should be measured against the reader'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.