Web Browsing Agent Sample

tl;dr

This is a sample Web Browsing Agent app that uses LangChain's OpenAIFunctionsAgent and Streamlit's StreamlitCallbackHandler.

You can try this out from this Sample App (OpenAI API key required).

demo.mp4

How it Works

• The agent uses custom-built DuckDuckGo's search tools and a page retrieval tool to perform web browsing to answer users' questions.
• It employs an agent that uses OpenAI Function Calling, ensuring stable operation.
• The agent operates as follows:
  • Considers search keywords based on the users' question
  • Executes the search using the considered keywords
    • The search results include titles, URLs, and snippets
  • Decides which page to view from the search results, retrieves the page content, and reviews the content
    • Writes the final answer if possible; otherwise, retrieves another page if more information is needed
• Of course, it is also possible to ask additional questions to the agent's answers.
• You can also check the details of the agent's operation.

Prerequisites

• You will need OpenAI's API_KEY

Details

Retrieving Page Content

• Utilizes the from_tiktoken_encoder function of RecursiveCharacterTextSplitter to fetch content in chunks of 1000 tokens.
  • If the agent wish to obtain more content, add a page parameter to fetch the page again
• As it's a prototype, caching has not been implemented
• It uses readability and html2text to extract the main body of the page
  • Note that html2text has a GPL3.0 license, so be cautious
• Some pages on the website load slowly at times, so we have also set up timeouts in requests.

Model

• Because the number of tokens used can significantly increase when a question and answer interaction occurs, using gpt-4 or gpt-3.5-turbo-16k is recommended.
  • You can select models in sidebar option buttons.

Reference

• Streamlit LLM Examples (Chat with Search)
  • I learned how to use StreamlitCallbackHandler from this code
• Build and Learn: AI App Development for Beginners: Unleashing ChatGPT API with LangChain & Streamlit
  • The code for the entire app was inspired by this book
  • Well, I wrote the book, so there's that haha 😎