Gemini's Function Calling Got Weird 🤷♂️ — Day 21: Building Voice AI Agents with Pipecat
Автор: John George
Загружено: 2025-02-08
Просмотров: 430
Описание:
A day of wrestling with Gemini's function calling inconsistencies and trying to make sense of unexpected prompt engineering challenges.
✓ Discovered weird behaviour with role vs task message handling
✓ Found inexplicable line break dependencies
✓ Attempted various temperature settings with no effect
✓ Hit potential token length issues (1,600 for consent, 2,500 for development)
✓ Tried Google AI Studio's suggestions (with mixed results)
Building an open-source voice AI agent framework. Following the real, unvarnished development journey through the challenges of working with LLMs, voice technology, and autonomous agents.
📦 Repository: https://github.com/askjohngeorge/pipe...
Connect with me:
🤝 / askjohngeorge
💻 https://github.com/askjohngeorge
🏠 https://askjohngeorge.com
0:00 Introduction
0:20 A tale of woe told through my git commits
1:48 Maddening behaviour with single line break difference
2:23 Max tokens and token length of node prompts
2:53 Started the day optimistic, adding examples of undesired outputs
3:14 Going around in circles
3:23 At wit's end, found Google AI Studio's prompt generator
4:03 Trying new prompting strategy
4:42 Next steps and wrap up
Повторяем попытку...
Доступные форматы для скачивания:
Скачать видео
-
Информация по загрузке:
