AI FAQ Agent with Multi-Turn Conversation

Overview

This project implements an AI-powered FAQ agent designed to answer user queries based on a provided set of frequently asked questions. It features a conversational Gradio web interface, allowing for multi-turn interactions. The agent utilizes a Retrieval Augmented Generation (RAG) pipeline with Langchain, HuggingFace sentence transformers for embeddings, Google Gemini for language generation, and ChromaDB as a vector store. It also includes logic for query complexity assessment and escalation to a human agent if needed.

Features

• Conversational UI: Interactive chat interface built with Gradio (gr.ChatInterface) supporting multi-turn dialogues.
• Retrieval Augmented Generation (RAG): Leverages Langchain to retrieve relevant FAQ snippets and generate context-aware answers.
• FAQ-Based: Answers are grounded in the content of data/faq_data.csv.
• Semantic Search: Uses HuggingFace sentence transformers (all-MiniLM-L6-v2) for creating text embeddings and ChromaDB for efficient similarity search.
• Powerful LLM: Utilizes Google's Gemini model for understanding queries and generating responses.
• Structured Output: Employs Pydantic models to ensure the LLM provides responses in a consistent format, including escalation flags.
• Escalation Logic: Can identify queries that are out-of-scope or require human intervention, providing appropriate escalation messages.
• Environment Configuration: Securely manages API keys using a .env file.
• Modular Design: Code is organized into logical modules (app.py, langchain_utils.py, utils.py).

Technology Stack

• Python 3.x
• Gradio: For the web interface.
• Langchain: Core framework for building the RAG pipeline and managing LLM interactions.
• HuggingFace Sentence Transformers: For generating text embeddings.
• Google Generative AI (Gemini): Language model for generation and understanding.
• ChromaDB: Vector database for storing and retrieving embeddings.
• Pandas: For loading and processing the FAQ data from CSV.
• Dotenv: For managing environment variables.

Setup Instructions

1. Clone the Repository (if applicable):

   # If this were a git repository:
   # git clone <repository_url>
   # cd <repository_directory>

   For now, ensure you have all project files in a single directory.

2. Create and Activate a Virtual Environment (recommended):

   python -m venv venv
   # On Windows
   .\venv\Scripts\activate
   # On macOS/Linux
   # source venv/bin/activate

3. Install Dependencies: Ensure your requirements.txt file is up-to-date with all necessary packages. Example packages include:

   gradio
   langchain
   langchain-community
   langchain-google-genai
   sentence-transformers
   chromadb
   pandas
   python-dotenv
   # Add other dependencies as needed
   

   Install them using pip:

   pip install -r requirements.txt

4. Set Up Environment Variables: Create a file named .env in the root directory of the project and add your Google API key:

   GOOGLE_API_KEY="YOUR_GOOGLE_API_KEY_HERE"

   Replace "YOUR_GOOGLE_API_KEY_HERE" with your actual API key.

Running the Application

1. Navigate to the project's root directory in your terminal.
2. Ensure your virtual environment is activated.
3. Run the Gradio application:

   python app.py

4. The terminal will display a local URL (e.g., http://127.0.0.1:7860). Open this URL in your web browser to interact with the AI FAQ Agent.

Project Structure

. (root directory)
├── .env                 # Environment variables (e.g., API keys) - Create this manually
├── app.py               # Main application script with Gradio interface and core logic
├── langchain_utils.py   # Utilities for Langchain RAG chain, prompts, and LLM setup
├── utils.py             # General utilities (e.g., document loading, embedding creation)
├── requirements.txt     # Python package dependencies
├── README.md            # This file
└── data/
    └── faq_data.csv     # CSV file containing questions and answers for the FAQ agent
└── rag_chain_tester.py  # Optional script for testing the RAG chain directly (CLI)

Testing

• Gradio UI: The primary way to test is by interacting with the chat interface launched by app.py. Try various types of questions:
  • Direct FAQ questions.
  • Questions requiring some interpretation.
  • Out-of-scope or irrelevant questions.
  • Conversational phrases (greetings, thanks).
  • Queries that might warrant escalation.
• Command Line Tester: You can use rag_chain_tester.py to test the RAG chain's responses directly in the terminal without the Gradio UI. This can be useful for debugging the core LLM and retrieval logic.

  python rag_chain_tester.py