AI Chatbot with LangChain and AWS

Social

Services used

Summary

This project develops an intelligent chatbot that generates automatic responses using generative artificial intelligence and a dynamic knowledge base. The solution integrates various Amazon Web Services (AWS) and leverages the LangChain libraries and the Wikipedia API to enrich responses.

Generative AI with RAG

A ReAct Agent has been implemented, utilizing the Retrieval Augmented Generation (RAG) technique to provide accurate and up-to-date responses. This agent interacts with two key tools:

• Knowledge Base Tool: we first create the knowledge base using Amazon Titan Embeddings as the embeddings model and Pinecone as the vector database in which the indexed embeddings will be stored. Then, we create the knowledge base tool using the create_retriever_tool method from LangChain transforming our retriever into a Tool object so that the agent can access it and execute it.

  from langchain.tools.retriever import create_retriever_tool
  
  retriever_tool = create_retriever_tool(
    name="Knowledge Base",
    description="With this tool you can access a custom knowledge base.",
    retriever=kb_retriever
  )

• Wikipedia Tool: using the same Tool class from before, and the from_function function, we create the Wikipedia tool from a custom function that checks first if the Wikipedia page exists, if it does, then the API summarizes the desired page.

  import wikipediaapi
  from langchain.agents.tools import Tool
  
  def search_wikipedia(self, title):
    wiki_wiki = wikipediaapi.Wikipedia('SampleProject/0.0 (example@example.com)', 'en')
    page = wiki_wiki.page(title)
        
    if page.exists():
      tools_logger.info(f"TOOLS: Page found:\n{page.summary}")
      return page.summary 
    else:
      tools_logger.info("TOOLS: Page not found in Wikipedia")
      return "No page was found on Wikipedia matching your search."
  
  
  wikipedia_tool = Tool.from_function(
    func=self.search_wikipedia,
    name="Wikipedia",
    description="With this tool you can access Wikipedia to summarize different topics that you don't know."
  )

AWS Services Used

The solution relies on several AWS services to ensure smooth integration and cost efficiency. The main services are:

• S3: Stores .pdf files containing structured information that the knowledge base tool uses to generate clear and precise responses.
• Lex: Provides the user interface for the chatbot, including voice services like Amazon Transcribe and Amazon Polly.
• Bedrock: Acts as the hub for foundational model (FM) selection and as the host for the knowledge base.
• Lambda: Functions as a serverless service facilitating calls to the AI agent and its tools, offering automatic scalability, security, and flexibility.

Libraries and Tools

• LangChain: Enables the integration and orchestration of different AI tools and knowledge bases, making the creation of a robust AI agent easier.
• Wikipedia-API: Provides access to extensive and updated data from Wikipedia, expanding the chatbot's capabilities.

Functionality Description

The general flow of the chatbot is as follows:

• Event Reception: AWS Lambda receives an event from Amazon Lex.
• Response Generation: Lambda calls the AI agent that uses LangChain to process the event.
• Knowledge Base Query: The agent employs RAG to search for information in the knowledge base hosted on Amazon Bedrock.
• Access to Wikipedia: If additional information is needed, the agent queries Wikipedia using its API.
• User Response: The generated response is formatted and sent back to Amazon Lex, which presents it to the user.

flowchart TD
    A(Lex sends an event) -->|event| B["Lambda creates and invokes 
    the agent using LangChain"]
    B --> C{"ReAct agent chooses
     and 
     returns the answer"}
    C -->|Direct answer| D["Lambda formats the response 
    and sends it back to Lex"]
    C -->|Knowledge base answer| D
    C -->|Wikipedia answer| D

    style A stroke:#168E64, fill:#168E64
    style B stroke:#BA661A, fill:#BA661A
    style C stroke:#434386, fill:#434386
    style D stroke:#BA661A, fill:#BA661A

Loading