Semantic Kernel - how to use HandleBars plans

Подписаться 1,3 тыс.

50% 1

In this video I show a different way to have Semantic Kernel solve a user question using a plan approach instead of standard ReAct framework. Using Plan technique is a different way to approach the same problem, letting a Large Language Model solve a user question using custom Agent available to the system.
At the beginning of the video, you can see Lucifer, he is a usual guest of my recording sessions, and sometimes he wants attention :).
▬ Contents of this video ▬▬▬▬▬▬▬▬▬▬
00:00 - Explanation of React Framework and its Limitations
01:15 - Introduction to Handlebars Planner in Semantic Kernel
02:20 - How to Use Handlebars Planner for Creating a Plan
03:06 - Benefits of Using Handlebars Planner
04:51 - Demonstration of Handlebars Planner in Action
07:01 - Importance of Plugin Description in Handlebars Planner
09:08 - Comparison of Handlebars Planner and React Approach
10:37 - Recap and Advantages of Using Handlebars Planner
12:44 - Conclusion and Closing Remarks

Наука

Опубликовано:

28 июл 2024

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 4

@James_PET 3 месяца назад

That is nicely explained. Thank you so much. A few follow-up questions: 1. if you are creating a generic chatbot that should be able to do "anything", let's say, to summarise a video, but also to write and send an email - do you create a HandleBars plan and give it to chatGPT to pick one? 2. What if you want to mix a few plans, from the example above -> Summarize this video and write and send email

@codewrecks 3 месяца назад

You will create agent for tasks, then it is GPT (or the LLM you want to use) that analyze the question of the user, looks at which agent it can use, and device a plan to solve user question. This lead to: a single user question will generate a unique plan. Plan is generated by LLM looking at the question of the user and the agent it can use.

@James_PET 3 месяца назад

@@codewrecks@codewrecks Thank you. I saw this in a few videos later, but thanks for clarifying. What would be super interesting is if you could record a video on how to save plans and call them later, especially if it is possible to have LLM pickup from pre-saved plans So, in your example, you have a plan for transcribing a video, if you have a plan that drafts an email on a specific topic and sends it via a plug-in. It would be amazing to see how to offer those plans to LLM so it can pickup depending on what user asks Thanks

@codewrecks 3 месяца назад

@@James_PET Plan are not meant to be saved, because they are based on user question. What you can do is create agents that aggregate other agents. I'll explain You have three agents: extract audio, extract text from audio, summarize timeline. Now if you ask "Can you summarize video XYZ.mp4" the agent will interrogate the LLM and based on three agents devise a plan. But you can do this: Create a fourth agent, call it "Summarize Video", and internally simply call the three previous agent (or the functions that the agents call). Now that agent is actually a plan agent that is capable of doing what you requested. I did not examined the possibility to save a plan directly, it could be interesting if a user want to save it. In this situation the scenario is User ask "Can you summarize video XYZ.mp4?" He/she got a plan in return, the plan is good and it press save. Give the plan a name and a description. Then the software should add that plan as agent to the list of available agents.