Amazon Nova Model Cookbook

Amazon Nova 2.0 models provide multimodal capabilities. The amazon.nova-2-lite-v1:0 model supports Text, Image, and Video for both input and output modalities. It features a 1M context window and supports up to 64K max output tokens. It is compatible with various document formats including pdf, csv, doc, docx, xls, xlsx, html, txt, and md.

Key features include support for the Converse API, InvokeAPI, Streaming, Batch Inference, Fine Tuning, Bedrock Knowledge Bases, Bedrock Agents, and Bedrock Prompt flows.

Model ID: amazon.nova-2-lite-v1:0
Input modalities: Text, Image, Video
Output Modalities: Text, Image, Video
Context Window: 1M
Max Output Tokens: 64K

The Citation-Aware Model Distillation Pipeline is a four-stage process designed to distill knowledge from large language models (LLMs) into smaller, specialized models specifically for citation-aware question answering. The goal is to create efficient models that maintain high-quality citation capabilities.

The pipeline follows these four stages:

1. Data Preparation: Preparing datasets (like SQuAD v2.0) with structured XML formats.
2. Model Distillation: Transferring knowledge from a teacher model to a student model.
3. Batch Inference: Running large-scale inference with robust error handling.
4. Evaluation: Assessing citation quality using specific metrics like coverage, correctness, and faithfulness.

Bedrock NotebookLM is a tool inspired by NotebookLM that uses AWS Bedrock LLMs and AWS Polly text-to-speech models to transform content into audio podcasts. It processes PDF and video files to generate natural, engaging dialogues and outputs the final result as an MP3 file.

Core Capabilities:

• PDF to Podcast: Convert uploaded PDF content into a podcast dialogue.
• S3 Integration: Process documents stored directly in Amazon S3.
• Video Processing: Upload video files to use as source content for the podcast.
• Gradio Interface: Provides a user-friendly web interface for interaction.

This pattern enables e-commerce customers to search for products using either images or text descriptions. By leveraging multimodal embeddings, the system can find visually and semantically similar items within a product catalog.

Supported Modalities:

• Images: Users can upload a photo to find similar products.
• Text: Users can type descriptions to find matching items.

Core AWS Architecture:

• Amazon Bedrock: Uses the Nova Multimodal Embedding model to generate embeddings for both image and text inputs.
• Amazon S3: Used for storing the product catalog.
• Vector Storage: Employs either Amazon S3 Vectors or Amazon OpenSearch Serverless as the vector database to perform similarity searches.

This sample demonstrates how to use Amazon Bedrock Nova Lite 2.0, a large language model (LLM) optimized for complex reasoning and agentic applications, with popular agent frameworks.

A primary implementation provided is an Academic Paper Analysis agent using the LangGraph ReAct pattern. This agent uses specific tools to:

• Load and parse PDF documents
• Summarize papers
• Extract research questions
• Extract key results and findings
• Identify research gaps and future work

Key technical features include LangGraph implementation of the ReAct pattern, PDF text extraction, and structured JSON outputs for research analysis.

The Email Assistant is a multi-agent system built using the Strands Agents SDK and a Streamlit interface. It leverages specialized agents to provide context-aware email assistance, including image generation, audio-based information retrieval, and professional composition.

Core Agents

• Image Agent: Generates images based on user descriptions and manages files.
• Audio RAG (Retrieval-Augmented Generation): Uses Bedrock Data Automation (BDA) to parse audio content (e.g., Amazon earnings calls) and answer queries based on that audio knowledge base.
• Email Assistant Agent: The primary orchestrator that uses web search, Audio RAG, and Image Generation to compose professional emails.
• Report Writing Agent: Specialized in systematic planning and report generation.

This tutorial series demonstrates production-ready patterns for integrating tools into Nova Act workflows using the @workflow decorator. Workflows defined here can be packaged into containers and deployed to the AgentCore runtime.

Key integration patterns covered include:

• Model Context Protocol (MCP): Standardizing tool discovery and invocation via servers (e.g., connecting to an AWS Documentation server).
• AgentCore Gateway: Centralized enterprise tool management, providing discovery, authentication, and governance.
• Security Guardrails: Implementing URL-based access controls and state guardrails to manage tool permissions.