AI Chatbot with Retrieval-Augmented Generation and Summarization

This project implements an AI-powered chatbot that answers finance news questions using retrieval-augmented generation (RAG), and advanced summarization with fine-tuned LLM adapters. It combines:

• Retrieval of relevant news documents
• Answer generation using large language models
• Summarization of answers or retrieved content

It also includes a Streamlit web app to interact with the chatbot.

---

Features

• Base LLM answering without context
• RAG with document retrieval
• Advanced RAG with improved answer quality
• Fine-tuned Summarizer adapters for:
  • Summarizing generated answers
  • Summarizing retrieved documents
• Streamlit Web UI for easy interaction

---

Repository Structure

├── ai_agent_source_code.py     # Main agent code (RAG + Summarization)
├── chatbot_web_app.py          # Streamlit web app for chatbot
├── Dockerfile                  # Docker container definition
├── lora_adapters.zip           # Fine-tuned LoRA weights for summarization
├── sample_output.pdf           # Example chatbot UI output
├── stitching_code.ipynb        # Jupyter Notebook for experimentation
├── unit_test.ipynb             # Notebook to test the end-to-end pipeline
└── README.md

---

Running with Docker

This repository includes a Dockerfile to help you containerize the full pipeline, including retrieval, generation, and summarization.

1. Prepare Your Data

• Download the finance news dataset from this link.
• Unzip it and place the finance_news folder in a convenient location.
• Download and unzip the LoRA adapter weights from this link.

2. Set Up Environment Variables

Create a .env file in your project root or mount it when running Docker:

PINECONE_API_KEY=your_pinecone_api_key_here
OPENAI_API_KEY=your_openai_api_key_here

3. Build the Docker Image

docker build -t genai_project .

4. Run the Container

Example Command:

docker run -p 8000:8000 \
  --mount type=bind,source="/path/to/finance_news",target=/home/finance_news \
  --mount type=bind,source="/path/to/lora_adapters",target=/home/lora_adapters \
  --env-file .env \
  genai_project

Results & Insights

• Base LLM → Outdated answers (knowledge cutoff October 2023)
• Base RAG → Better coverage with retrieved documents
• Advanced RAG (no fine-tuning) → More detailed and relevant answers with fewer hallucinations
• Advanced RAG + Summarizer (on generated answer) → Produces clear, concise summaries of generated answers
• Advanced RAG + Summarizer (on retrieved docs) → Summarizes retrieved documents before generation, reducing noise and improving conciseness

Preferred approach: Summarizing retrieved documents first generally yields the most concise and relevant answers.

---

Credits

Project developed by Glenys Charity Lion.

---

Acknowledgments

This project was developed as part of a Generative AI course. I appreciate your interest and welcome feedback!