Smart Research Assistant Agent

An autonomous AI agent powered by Amazon Nova 2 Lite that reasons, selects tools, reads PDFs, and synthesizes intelligent research answers — all without human intervention. #AmazonNova

Comment

Inspiration

As a student exploring AI, I was fascinated by how modern AI agents can do more than answer questions — they can actually think, plan, and take actions. I wanted to build something that demonstrates this "agentic" behavior using Amazon Nova, showing how a foundation model can power an autonomous research assistant that reasons through problems step by step.

What it does

Smart Research Assistant Agent is an autonomous AI agent that helps users research topics and analyze documents. When you ask a question, the agent:

  1. Reasons about your question using Amazon Nova 2 Lite
  2. Decides which tool to use (web search, PDF analysis, or data analysis)
  3. Executes the selected tool to gather real information
  4. Synthesizes a comprehensive answer using Nova's reasoning

The standout feature is real PDF document analysis — the agent can read actual PDF files using PyPDF2, extract text from every page, and use Nova to provide intelligent analysis of the content.

How I built it

  • Amazon Nova 2 Lite (amazon.nova-lite-v1:0) via AWS Bedrock Converse API powers all reasoning and synthesis
  • Python for the agent framework with a modular tool system
  • PyPDF2 for real PDF file reading and text extraction
  • boto3 for AWS Bedrock integration
  • The agent uses a structured THOUGHT → ACTION → INPUT loop where Nova decides which tool to use, then synthesizes the results into a final answer

Challenges I ran into

  • Getting the Bedrock Converse API message format right — the content field must be a list of content blocks ([{"text": "..."}]), not a plain string
  • Handling API throttling gracefully with retry logic when hitting daily token quotas
  • Making the action parser robust enough to handle different LLM output formats (Nova might return ACTION: analyze_pdf("file.pdf") or ACTION: ANALYZE_PDF)
  • Handling edge cases in PDF extraction where some pages return None from extract_text()

Accomplishments that I'm proud of

  • Built a true agentic system that reasons and acts autonomously, not just a chatbot wrapper
  • The PDF analysis feature uses real file I/O — not simulated — making it genuinely useful for research
  • Clean, well-documented code with lazy initialization, error handling, and robust parsing
  • The agent architecture is extensible — adding new tools is as simple as adding a method

What I learned

  • How the AWS Bedrock Converse API works and how to integrate Amazon Nova models
  • The difference between a chatbot (prompt → response) and an agent (reason → plan → act → synthesize)
  • How to build a tool-use framework where an LLM decides which tools to call
  • The importance of error handling and retry logic in production AI systems

What's next for Smart Research Assistant Agent

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

Built with

Python, Amazon Nova 2 Lite, AWS Bedrock, PyPDF2, boto3

Built With

  • amazon-nova-2-lite
  • aws-bedrock
  • boto3
  • python
Share this project:

Updates