Smart Research Assistant Agent

An Autonomous AI Agent Powered by Amazon Nova 2 Lite #AmazonNova

Amazon Nova AI Hackathon 2026 | Category: Agentic AI | Built by Devika

---

What Is This Project?

This is a Smart Research Assistant Agent — an autonomous AI system that can think, plan, and act on its own to answer your research questions. Unlike a regular chatbot that just responds to what you type, this agent actually reasons about your question, decides which tool to use, executes that tool, and then synthesizes a clear answer — all powered by Amazon Nova 2 Lite on AWS Bedrock.

Think of it like having a research assistant who can read your PDF documents, search for information, analyze data, and give you well-thought-out answers — all automatically.

---

How It Works (Agent Architecture)

┌─────────────┐      ┌──────────────────┐     ┌────────────────┐      ┌───────────┐      ┌──────────────┐
│  User Query │ ──▶ │ Nova Reasoning   │ ──▶ │ Tool Selection │ ──▶ │ Execution │ ──▶  |  Synthesis   │
│  (Question) │     │  (Amazon Nova 2) │      │  (Which tool?) │      │ (Run it!) │      │ (Answer!)    │
└─────────────┘     └──────────────────┘      └────────────────┘      └───────────┘      └──────────────┘

The agent follows a 5-step reasoning loop for every question:

1. Receive — The user asks a question in plain English
2. Reason — Amazon Nova 2 Lite analyzes the question and thinks about the best approach
3. Decide — The agent picks the right tool (web search, PDF reader, or data analyzer)
4. Execute — The selected tool runs and gathers real information
5. Synthesize — Nova 2 Lite processes everything and produces a comprehensive answer

This is what makes it an agent — it doesn't just respond, it thinks and acts.

---

Key Features

PDF Document Analysis (The Highlight!)

The agent can read real PDF files using PyPDF2 — not simulated, actual file I/O. It extracts text from every page, handles edge cases (blank pages, scanned documents), and sends the content to Nova for intelligent analysis.

agent.run("What are the key findings in research_paper.pdf?")
agent.run("Summarize the methodology in thesis.pdf")
agent.run("Compare results in paper1.pdf vs paper2.pdf")

Web Research

Ask any research question and the agent will search for information and synthesize an answer.

agent.run("What is the current state of agentic AI systems?")

Data Analysis

The agent can perform market analysis, sentiment analysis, and more.

agent.run("What does market analysis say about AI growth?")

Interactive Terminal Interface

Run the agent and interact with it in real-time. Type any question, or type demo for a guided demonstration.

---

How Amazon Nova Is Used

Amazon Nova 2 Lite (amazon.nova-lite-v1:0) serves as the brain of the entire agent. It's used in two critical places:

1. Reasoning Step — Nova analyzes the user's question and decides which tool to use. It outputs a structured THOUGHT → ACTION → INPUT plan.

2. Synthesis Step — After a tool returns results (e.g., text extracted from a PDF), Nova takes that raw information and produces a clear, comprehensive answer.

The agent communicates with Nova through the AWS Bedrock Converse API, which provides a clean interface for sending prompts and receiving responses.

---

Quick Start

Prerequisites

• Python 3.8 or higher
• An AWS account with Amazon Bedrock access enabled
• Access to the Amazon Nova 2 Lite model in us-east-1 region

Step 1: Clone the Repository

git clone https://github.com/YOUR_USERNAME/amazon-nova-hackathon.git
cd amazon-nova-hackathon

Step 2: Install Dependencies

pip install -r requirements.txt

Step 3: Configure AWS Credentials

# Option A: Use AWS CLI
aws configure
# Enter your Access Key ID, Secret Access Key, and region (us-east-1)

# Option B: Set environment variables
export AWS_ACCESS_KEY_ID=your_access_key
export AWS_SECRET_ACCESS_KEY=your_secret_key
export AWS_DEFAULT_REGION=us-east-1

Step 4: Run the Agent

python amazon_nova_project.py

You'll see a welcome screen. Type any research question, or type demo to see the agent in action!

Step 5: Try PDF Analysis

Place any PDF file in the same folder, then:

python amazon_nova_project.py
# Then type: Analyze the key findings in sample.pdf

Or run the interactive PDF examples:

python pdf_analysis_examples.py

---

Project Structure

amazon-nova-hackathon/
├── amazon_nova_project.py      # Main agent code (the core of the project)
├── pdf_analysis_examples.py    # 7 interactive PDF testing examples
├── requirements.txt            # Python dependencies
├── sample.pdf                  # Sample PDF for testing
├── demo.html                   # Interactive HTML demo of the agent
├── DEVPOST_SUBMISSION.md       # Devpost submission text (copy-paste ready)
└── README.md                   # This file

---

File Descriptions

File	What It Does
amazon_nova_project.py	The main agent — contains ResearchTools (4 tools including PDF analysis), ResearchAgent (reasoning loop, tool selection, Nova API calls), and interactive CLI
pdf_analysis_examples.py	7 runnable examples showing different ways to use PDF analysis: basic analysis, information extraction, PDF comparison, Q&A, batch processing, research workflows, and custom queries
requirements.txt	Lists all Python packages needed: boto3, botocore, python-dotenv, PyPDF2
sample.pdf	A sample research paper PDF for testing the PDF analysis feature
demo.html	An animated HTML demo that shows how the agent works — open in any browser

---

Technical Details

Tools Available to the Agent

Tool	Description	Real/Simulated
analyze_pdf	Reads PDF files, extracts text from all pages, handles edge cases	Real (PyPDF2)
search_web	Searches for information on a topic	Simulated (placeholder for real API)
fetch_document	Retrieves document content	Simulated
analyze_data	Performs data analysis (market, sentiment)	Simulated

Amazon Nova API Integration

The agent uses the Bedrock Converse API to communicate with Nova:

client = boto3.client('bedrock-runtime', region_name='us-east-1')
response = client.converse(
    modelId="amazon.nova-lite-v1:0",
    messages=[{
        "role": "user",
        "content": [{"text": prompt}]  # Content must be a list of blocks
    }],
    inferenceConfig={
        "maxTokens": 1024,
        "temperature": 0.7
    }
)

Key Design Decisions

• Lazy client initialization — The AWS Bedrock client is only created when the first API call is made, not at import time. This means you can import the module and test PDF extraction even without AWS credentials configured.
• Robust action parsing — The LLM might return ACTION: analyze_pdf("file.pdf") or ACTION: SEARCH_WEB. The parser handles all variations using regex cleanup and fuzzy matching.
• Safe PDF handling — Handles None from extract_text() (blank/scanned pages), truncates very long documents to stay within token limits, and provides clear error messages.

---

Demo

Open demo.html in your browser to see an interactive animation of the agent processing three different types of queries:

1. Web Research — Asking about agentic AI trends
2. PDF Analysis — Reading and analyzing a research paper
3. Data Analysis — Market analysis of AI growth

The demo includes an architecture diagram showing the agent's 5-step reasoning loop.

---

Tech Stack

• AI Model: Amazon Nova 2 Lite (amazon.nova-lite-v1:0)
• Cloud Platform: AWS Bedrock (Converse API)
• Language: Python 3
• PDF Processing: PyPDF2
• AWS SDK: boto3

---

What Makes This Project Special

Aspect	Why It Matters
True Agentic Behavior	The agent reasons, plans, selects tools, and synthesizes — not just prompt-in, text-out
Real PDF Analysis	Actual file I/O with PyPDF2, not simulated responses
Nova as the Brain	Amazon Nova 2 Lite drives both the reasoning and synthesis steps
Production-Ready Patterns	Lazy initialization, error handling, input sanitization
Extensible Design	Easy to add new tools — just add a method to ResearchTools and register it

---

Future Improvements

• Connect search_web to a real search API (Google, DuckDuckGo, or Bing)
• Add image analysis using Amazon Nova's multimodal capabilities
• Add conversation memory so the agent remembers previous questions
• Support for more document formats (Word, Excel, HTML)
• Deploy as a web application with a proper UI

---

Built With

Amazon Nova 2 Lite on AWS Bedrock | Python | PyPDF2 | boto3

#AmazonNova | Amazon Nova AI Hackathon 2026 | Agentic AI Category

---

Built by Devika