I have recently released SumMario, an AI agent based on Chatgpt which is capable of searching the internet for news, summarizing the news, summarizing Youtube video contents, and read txt, ppt, pdf files. You can find the codes on GitHub.
SumMario also acts as a personal assistant, being able to access its memory and keep conversating with the user about the stored data.
Here is a short demo of it for extracting daily news from the web and chatting about the obtained data.
How Does It Work?
SumMario relies on the lastest LLM technology like ChatGPT. It is a web-based agent, developed using Gradio, and LangChain for the integration of GPT agents.
SumMario is made of 3 different agents: SuperAgent, DataRetriever, and ConversationalAgent.
The SuperAgent is the higher level AI agent responsible for managing the DataRetriever and the ConversationalAgent. Basically, it uses the other two agents as tools and, given specific prompts as system message and the user message, it is able to select which agent to choose to complete a task.
The ConversationalAgent is a basic Chatbot based on GPT. It’s purpose is to use directly LLM knowledge and reasoning skills to complete the user’s requirements. It has no access to external tools but only to its stored memory.
The DataRetriever, instead, is an AI agent with access to tools. The tools include GoogleWebSearch, WebPageSummarizer, VideoSummarizer, FileReader. Given the user’s query, the DataRetirever is thus able to select the proper tool and extract text from the source, it being a web page, a Youtube video, or a local file (pdf, ppt, word).
Each tool of the DataRetriever stores extracted text from the sources in a local temporary txt file. This file is accessible to the SuperAgent too, which, on instruction of the SuperAgent, can open the temporary txt file and help the DataRetriever complete the user’s defined task.
Francesco Cursi 