Data Overview for This Category
This category focuses on security training and phishing simulations for code and change review scenarios. Daily documents include training slides, simulation plans, phishing email templates, and simulation result reports. Some data is linked to code commit records and review logs. Data comes from employee directories, email gateways, training platforms, and click statistics. Typical users include security operations teams, internal corporate training departments, and external security service providers. Data update frequency is usually quarterly or annually. Core fields include simulation audience scope, email delivery records, employee click feedback, and rectification verification results. Documents are interdependent, requiring matching of corresponding audiences and statistical data.
Constraints Imposed on Multi-turn Dialogue and Prompt Engineering
This category focuses on security training and phishing simulations for code and change review scenarios. Decentralized data sources and low update frequency mean multi-turn dialogue must adapt to non-real-time static datasets. Prompts must clearly distinguish between static rules and dynamic interactions. Audience screening for simulations relies on employee directory fields. Prompts must include field validation logic to avoid calling unauthorized code review-related data. Multi-turn interactions during email template design must support dynamic variable replacement for code repositories and commit records. Prompts must clearly define variable formats and value ranges. Dialogues during result review must link historical simulation and code review data. Historical reports and review log fields must be passed via variables to ensure coherent context. Since errors directly relate to security risks, multi-turn dialogue must add compliance validation steps to avoid generating simulation content that fails audit requirements.
Configuration Guidelines
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Approach | Rationale |
|---|---|---|
variables | Pass fields such as simulation ID, employee department, code repository path | Adapt to static data association requirements in code and change review scenarios. Pass preset personnel and code-related data via variables to avoid repeated calls to decentralized data sources |
messages | Concatenate system prompts and user input by conversation turn, retain the first 3 rounds of context | Low data update frequency for this category means long context is unnecessary. Retaining core interaction turns reduces token consumption while ensuring coherent dialogue logic related to code reviews |
chatId | Bind the current simulation conversation ID | Distinguish conversation contexts for different simulations, avoid cross-session data confusion, and ensure isolation of code review and security training data |
| Maximum context length | 8000–12000 characters | Match the average length of documents such as phishing email templates, simulation reports, and code review logs, ensuring prompts can fully load core rules and historical data |
| Conversation timeout | 600 seconds | Adapt to long-duration interaction scenarios such as simulation plan design and review analysis, preventing loss of code review-related progress due to mid-session interruptions |
| Variable validation toggle | Enabled | Prevent passing undefined variables fields, avoid generating simulation content that fails audit requirements, and align with compliance requirements for this category |
Three Common Misconfigurations
- Symptom: After passing
variablesvia API, prompt variables are not replaced, and returned content is unrelated to preset phishing email scenarios related to code reviews. Cause: Variable placeholders are not declared using the{{variable name}}format in the system prompt, or the key names of passedvariablesfields do not match the placeholders. - Symptom: Extraction of user-input audience information related to code review simulations fails in multi-turn dialogue, and the specified reply is triggered without waiting for user supplementary input. Cause: Conversation interruption waiting logic is not configured, or text extraction rules do not cover common expression formats for simulation audiences.
- Symptom: Using an English prompt to request the model to output English content, but returned content is only in Chinese. Cause: The system prompt does not clearly specify output language priority, or language format constraints are not enforced in the system message within
messages.
How to Verify Successful Configuration
- Manually construct an API request that includes
variables, check whether returned content correctly replaces preset values such as code repository paths and employee departments. - Trigger test scenarios where text content extraction fails, confirm the system returns a preset prompt and waits for user supplementary input.
- Configure English system prompts and English knowledge bases, verify that the language format of model returned content meets requirements.
- View conversation logs, confirm that
chatId,messages, andvariablesparameters are correctly recorded and passed.
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.