Gemini AI & Prompt Engineering Skill

This skill governs all AI-related logic within the TalentFilter platform, specifically focusing on the AIService in the backend.

1. Google Gemini API Integration

Model Versioning: Always use the most capable version (gemini-2.5-flash).
Safety Settings: Configure AI safety settings via the google-genai SDK to ensure the AI doesn't refuse valid recruitment/screening tasks while maintaining professional boundaries.
Async Execution: AI calls are high-latency; always use await and ensure they are proxied through the FastAPI backend to avoid blocking the main thread.

Strict Schema: Every AI call must use a Pydantic model to define its response structure.
JSON Request: Set the response_mime_type to application/json in the Gemini config.
Validation: Always wrap the AI parsing logic in a try-except block. If parsing fails, implement a retry mechanism or return a structured error response.
Language Awareness: Every prompt must explicitly specify the output language based on the user's locale.
Dynamic Scope: Avoid hardcoding question counts. Instruct the AI to determine the optimal depth (e.g., "Generate as many questions as needed to thoroughly vet this specific role").