Valid Schema, Wrong Content: Using Jev to Guard a Google ADK Agent

Описание и применение

Схема‑валидный JSON, но не соответствует содержанию: интеграция Jev в агент ADK позволяет автоматически проверять и корректировать подписи видео‑сегментов, чтобы они вписывались в заданную длительность.

Задача

Schema‑valid JSON produced by a video‑generation agent can still contain content that does not fit the specified duration, leading to rushed narration and mid‑sentence cuts.

Результат

The workflow demonstrates that schema‑validation alone is insufficient; integrating a content‑fit tool (Jev) within an ADK agent ensures captions match segment durations, producing a polished short video.

Рабочий процесс

• 1. Define a VideoTimeline schema for the storyboard.
• 2. Use Google ADK to create a director_agent with Gemini‑3.1‑pro‑preview model and output_schema set to VideoTimeline.
• 3. Add a FunctionTool that calls Jev’s check_segment to evaluate caption fit.
• 4. In the agent’s instruction, ask it to call check_segment on every segment and regenerate any flagged as needs_review, retrying up to twice.
• 5. Run the agent to generate the storyboard; Jev returns confidence scores and flags.
• 6. Regenerate flagged segments until all captions fit their durations.
• 7. Stitch the validated storyboard into a video.
• 8. Verify the final video has smooth narration and no mid‑sentence cuts.

Доказательства

Schema-valid is not content-correct – schema validation doesn’t catch caption length issues.

Источник

Wrap Jev as an ADK tool the agent uses to self‑check.

Источник

Pairing output_schema with tools on the same Agent is officially supported by ADK.

Источник