Amazon Bedrock

Amazon Bedrock excels in situations where organizations require rapid access to advanced AI functionalities without the complexities of data preparation, model development, or infrastructure management. This service supports a variety of use cases, such as creative content generation, dialogue system construction, text summarization, multilingual text generation, and sophisticated image generation tasks.

Before utilizing models, users must accept the terms of third-party marketplaces, excluding Amazon Titan Models. Notable use cases facilitated by Bedrock include:

• Text generation
• Search capabilities
• Image creation
• Chatbots
• Text summarization
• Personalization

Sagemaker Jumpstart

Sagemaker Jumpstart is designed to handle a wider array of machine learning tasks that demand in-depth control over model creation, training, and deployment. It is suited for tasks such as predictive analytics, recommendation systems, anomaly detection, and any other scenarios requiring tailored machine learning models. Users enjoy complete control over compute configurations and data preparation.

Jumpstart provides access to a model playground for experimentation (access needs to be requested) and offers a variety of capabilities across Foundation Models, Computer Vision models, and NLP models:

• A hub featuring built-in algorithms and pre-trained foundational models.
• Pre-configured training and inference scripts for convenience.
• An intuitive interface along with a Python SDK-based workflow.
• Interactive notebooks with practical examples to aid your work.

Amazon Q

Amazon Q is specifically crafted to function as a chat assistant within organizations. It aids in workplace management, code generation and transformations, generative business intelligence capabilities, conversational AI, and enhancing enterprise productivity. By utilizing its generative-AI features, Amazon Q facilitates task streamlining, relevant information provision, and accelerated problem-solving. It integrates closely with Amazon Code Whisperer (for code transformation) and Amazon QuickSight (for generative BI).

Amazon Bedrock

Amazon Bedrock is a fully managed service that equips organizations with advanced AI capabilities without necessitating data preparation, model development, or infrastructure management. It offers high-performing foundational models (FMs) through a single API, enabling the creation of generative AI applications while simplifying development with a focus on security, privacy, and responsible AI. There are no servers or infrastructure to manage as it operates in a serverless environment.

Sagemaker Jumpstart

Conversely, Sagemaker Jumpstart comprises a comprehensive suite of functionalities aimed at a broader spectrum of machine learning tasks. It allows users to prepare, build, train, and deploy high-quality ML models swiftly. Users can select from a variety of machine learning models and frameworks, including foundational models like Falcon-40b and Llama 2, allowing for selection tailored to specific needs. Sagemaker Jumpstart manages the underlying compute and provides an HTTP endpoint that can be invoked from the user's code.

Amazon Q

Lastly, Amazon Q serves as a fully managed, generative-AI powered chat assistant. It can be configured to respond to inquiries, summarize information, generate content, and execute tasks based on organizational data. Built on Amazon Bedrock's LLMs, Amazon Q enhances task efficiency, expedites problem resolution, and delivers immediate, pertinent information to users. It features an intuitive interface for configuring and utilizing the chat assistant. Amazon Q can be accessed via the AWS Console, Amazon Q API, CLI, and SDKs.

Amazon Bedrock

The Amazon Titan Multimodal Embeddings foundation model in Amazon Bedrock possesses robust multimodal capabilities. It can generate embeddings for text, images, and combinations thereof, which can be utilized to enhance multimodal search and recommendation experiences. For instance, users could search for images using descriptive phrases, utilize images as queries, or input both text and images together. The model comprehends the relationships between images and text, yielding more precise and contextual search results. Amazon Titan Multimodal Embeddings is currently available in Amazon Bedrock and can be customized through methods like fine-tuning to better understand a customer’s unique content.

Sagemaker Jumpstart

The Amazon SageMaker JumpStart Industry SDK enables users to access common public financial documents containing both text and tabular data. It facilitates processing and extracting features from text fields in documents such as SEC filings.

The corporate credit scoring solution exemplifies how to create a model that incorporates text data from SEC filings alongside standard numerical financial ratios for prediction purposes.

Pre-trained language models offered through JumpStart have been trained on extensive financial text corpora, allowing them to comprehend relationships between language, numbers, and tables presented in documents. Example notebooks illustrate how to construct models that combine text features with standard tabular data for tasks such as classification and regression. JumpStart aims to simplify the initiation of multimodal ML workflows, enabling users to leverage existing tools and models or fine-tune models based on their mixed data. The objective is to help organizations extract greater insights by merging language and standard data within their ML applications.

Amazon Q

Currently, Amazon Q does not feature a specific machine learning model with multimodal capabilities. As an AI assistant, its design emphasizes helpfulness and clarity in natural language conversations. It utilizes multiple foundational models from Amazon Bedrock to interpret language, provide relevant responses, and execute tasks.

Amazon Bedrock

Users interact with Amazon Bedrock through a simple API that invokes a foundational model. Pricing for Amazon Bedrock is determined by the quantity of input and output tokens processed, rather than the underlying computational capacity. Two types of inference options are available for invocation.

Sagemaker Jumpstart

Sagemaker Jumpstart provides an HTTP endpoint that users can access from their code. The underlying compute is managed by Sagemaker, permitting users to select the appropriate infrastructure for training and deploying their models. Sagemaker Jumpstart facilitates the complete lifecycle of ML tasks, including various inference options available for deploying LLMs as endpoints.

Amazon Q

Amazon Q can be accessed through the AWS Console, Amazon Q API, CLI, and SDKs. The chat assistant offers an interactive chat application for end users, leveraging enterprise data and extensive knowledge from large language models to provide relevant information and perform tasks. Amazon Q can connect to over 27 data connectors, allowing it to ingest document content and convert it into vector embeddings, setting the stage for personalized interactions using that data.

Amazon Bedrock

With Amazon Bedrock, users can customize foundational models through ongoing pre-training (using unlabeled data) and fine-tuning (using labeled data). This capability allows users to tailor the models to their specific needs and applications.

Sagemaker Jumpstart

Sagemaker Jumpstart provides extensive customization options. Users can employ their own algorithms, built-in options, or choose from a vast selection of models available on the AWS Marketplace or from third-party sources. This level of customizability enables users to fine-tune models to meet specific demands. Fine-tuning may be beneficial for:

• Adapting models to specific business needs
• Ensuring models effectively handle domain-specific language, such as technical terms or industry jargon
• Improving performance for particular tasks
• Generating accurate, context-aware responses in applications
• Producing factual, less biased, and more tailored outputs

There are two types of fine-tuning available in Sagemaker Jumpstart:

• Domain Adaptation Fine-tuning: This process allows leveraging pre-trained foundational models and adapting them to specific tasks using limited domain-specific data, modifying the model's weights accordingly.
• Instruction-based Fine-tuning: This method uses labeled examples formatted as prompt-response pairs to enhance the performance of pre-trained foundational models on specific tasks. Fine-tuned Language Net (FLAN) models employ instruction tuning to improve adaptability for solving various downstream NLP tasks.

Amazon Q

As a managed service, Amazon Q does not provide model customization options. It is built on Amazon Bedrock's LLMs and offers functionality that is a subset of one of the models available within the Bedrock ecosystem.

Amazon Bedrock

Amazon Bedrock simplifies the setup process by allowing users to select a suitable pre-trained foundational model and customize it with their data. Users do not need to manage infrastructure, as that is taken care of by AWS. They can easily build applications by choosing a foundational model, fine-tuning when necessary, and using Bedrock APIs to submit prompts and receive responses.

Sagemaker Jumpstart

Sagemaker Jumpstart supports the entire lifecycle of machine learning, from data labeling to model deployment and monitoring. Users prepare data, select or create algorithms, train models, and deploy them. Infrastructure management is integrated into the ML lifecycle, which requires a more substantial effort. Users must prepare data, select or create algorithms, train models, and subsequently deploy them via a pipeline.

Amazon Q

As a managed service, Amazon Q streamlines the ML lifecycle by offering a fully managed, generative-AI powered chat assistant. Users can set up Amazon Q using the AWS Console, CLI, or SDKs, without needing extensive management of the ML lifecycle.

Amazon Bedrock

Amazon Bedrock processes user data within the AWS environment, ensuring data protection and security. It guarantees that user data is not used to train the foundational models. All data is encrypted and remains within the user's Virtual Private Cloud (VPC). Organizations with stringent security requirements should assess these considerations thoroughly.

Sagemaker Jumpstart

With users retaining control over infrastructure, Sagemaker Jumpstart allows for complete oversight of data protection and security. Users can encrypt data both at rest and in transit, manage data access via Identity and Access Management (IAM) roles, and comply with regulations through AWS's robust compliance measures. Sagemaker Jumpstart can be utilized within a Virtual Private Cloud (VPC) for network-level control.

Amazon Q

Similar to Amazon Bedrock, Amazon Q processes data within the AWS environment. It ensures data protection through encryption during data transit and at rest, offers cloud trail logs for monitoring, and manages data access via IAM roles. Additionally, Amazon Q incorporates a retrieval augmented generation (RAG) system to acquire relevant data from company data sources and enrich prompts.

Amazon Bedrock

Amazon Bedrock supports leveraging knowledge bases to provide contextual information from a company’s private data sources. This functionality enables retrieval augmented generation (RAG) to yield more relevant, precise, and tailored responses. Bedrock automatically retrieves documents, converts them into embeddings, and stores them in a vector database.

The Knowledge Bases feature for Amazon Bedrock is a fully managed capability that facilitates the entire RAG workflow from data ingestion to retrieval and prompt augmentation, eliminating the need to construct custom integrations with data sources and manage data flows. Built-in session context management allows applications to support multi-turn dialogues seamlessly. Users can simply specify the location of their data in Amazon S3, and Knowledge Bases for Amazon Bedrock will automatically retrieve the documents, segment them into text blocks, convert the text into embeddings, and store those embeddings in the vector database.

For those without an existing vector database, Amazon Bedrock will create an Amazon OpenSearch Serverless vector store. Users can also specify an existing vector store from supported databases, such as Amazon OpenSearch Serverless, Amazon Aurora, Pinecone, and Redis Enterprise Cloud, with future support for MongoDB.

Sagemaker Jumpstart

Users can implement Retrieval Augmented Generation (RAG) with LLMs regardless of whether they use Bedrock or Jumpstart. RAG retrieves information from outside the language model (non-parametric) and enhances prompts by incorporating the relevant retrieved data into context. With Sagemaker Jumpstart, users can deploy their chosen LLM. For example, they can utilize two SageMaker endpoints for the LLM (Flan T5 XXL) and embedding model (GPT-J 6B), employing an in-memory FAISS vector database. Users can also utilize both native (Open Search) and third-party vector databases (Chroma DB, FAISS) that support vector embeddings and similarity searches. The Langchain Library aids in handling large documents and segmenting them into smaller parts.

Sagemaker Jumpstart also offers additional features, including support for custom data sources like Redshift Clusters, Snowflake, and Databricks. It includes data processing capabilities with Apache Spark and provides tools for data wrangling and streaming custom data.

Amazon Q

Amazon Q comes equipped with an integrated RAG system and vector database. A RAG model consists of two components:

• A retrieval component that retrieves relevant documents based on the user query.
• A generation component that processes the query alongside the retrieved documents to generate a response using a large language model.

A data source serves as a document repository. If there isn't a queryable interface for data retrieval, Amazon Q creates an index that can synchronize with a data source, enabling its use in chat applications. This index can crawl and synchronize documents from the data source into an Amazon Q index at regular intervals. The data source connector enables integration and synchronization of data from multiple repositories into a single container. Amazon Q supports several data connectors, allowing users to build their generative AI solutions with minimal configuration.

Additionally, Amazon Q features a plugins capability that allows integration with third-party services such as Jira and Salesforce, enabling users to perform specific actions within the Amazon Q chat interface, such as creating tickets.

Amazon Bedrock

Pricing for Amazon Bedrock is usage-based, with costs incurred for the number of tokens processed by the foundational models and the computational time utilized for fine-tuning. A token represents a basic unit of text, consisting of a few characters. For image generation models, fees apply for each image produced. The cost structure scales with user requirements, offering two pricing plans: On-Demand and Batch, or Provisioned Throughput.

• On-Demand: Users pay solely for what they utilize, with no time-based commitments. For text generation models, charges apply for each input token processed and each output token generated. For embedding models, fees are incurred for each input token processed.
• Batch: Users can submit a set of prompts as a single input file and receive responses as a single output file, facilitating large-scale simultaneous predictions. Responses are processed and stored in the user's Amazon S3 bucket for future access. Batch mode pricing aligns with On-Demand pricing.
• Provisioned Throughput: This mode is designed for large, consistent inference workloads that require guaranteed throughput. Custom models are accessible exclusively through Provisioned Throughput. A model unit indicates a specific throughput, measured by the maximum number of input or output tokens processed per minute. Provisioned Throughput pricing is charged hourly, allowing flexibility in choosing between one-month or six-month commitment terms.

For detailed pricing information, please refer to Amazon Bedrock Pricing.

Sagemaker Jumpstart

Costs for Sagemaker Jumpstart depend on the infrastructure utilized for model training and inference throughout the duration. Unlike Bedrock API calls, it is not a pay-as-you-go model. For serverless inference, costs are proportional to inference duration. Sagemaker Jumpstart presents two payment options: On-Demand Pricing and SageMaker Savings Plans. On-Demand Pricing allows users to pay as they go without any minimum fees or upfront commitments. SageMaker Savings Plans provide a flexible, usage-based pricing structure in exchange for a commitment to consistent usage.

Costs associated with Sagemaker can be significant for inference and training. Various inference techniques, such as Real-Time Inference, Asynchronous Inference, and Serverless Inference, incur costs that are directly proportional to duration and compute specifications. Extended durations for training or inference with high compute specifications will result in higher costs.

Separate charges apply for training and data preparation.

For accurate pricing details, please consult Sagemaker Pricing.

Amazon Q

Amazon Q presents two pricing options: Amazon Q Business and Amazon Q Builder. Amazon Q Business is priced at $20 per month per user, while Amazon Q Builder costs $25 per month per user. Additional charges may apply for document indexing. The Amazon Q index incurs a fee for units of 20K documents at a rate of $0.14/hour ($100/month) to support business knowledge needs. Users can start with one unit and add more storage units as necessary, with each unit including 100 hours of connector usage monthly.

For further pricing details, please see Amazon Q pricing.

Amazon Bedrock

Amazon Bedrock is available across multiple AWS regions globally, including the US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), Europe (Frankfurt), and AWS GovCloud (US-West). Certain features, such as model evaluation, agents, and knowledge base, are currently limited to US regions (as of the time of writing). Amazon Bedrock provides a unified API for experimenting with and customizing high-performing foundational models from leading AI companies, streamlining the development of complex generative AI applications across AWS regions.

Sagemaker Jumpstart

SageMaker is accessible across most commercial AWS regions, encompassing regions in the Americas, Europe, Middle East, Africa, Asia Pacific, India, and AWS GovCloud. Availability has expanded over time, with recent additions including Osaka, Jakarta, and Calgary. Accessibility for some SageMaker features varies—certain functionalities, such as model evaluation and deployment, may initially be limited to specific regions before broader expansion. The comprehensive regional coverage enables customers to develop and deploy machine learning models using AWS infrastructure closest to their users or data sources.

Amazon Q

Amazon Q is currently offered in a limited preview release across two AWS regions—US East (N. Virginia) and US West (Oregon). Data from interactions with Amazon Q is stored in these US regions to ensure privacy and security. Designed to be helpful, harmless, and honest, Amazon Q employs the global AWS infrastructure of regions and availability zones, which are highly available and resilient, ensuring automatic failover between zones without service interruption.

Amazon Bedrock

Amazon Bedrock presents a moderate learning curve due to its focus on working with foundational models and developing generative AI applications. Comprehensive documentation, tutorials, and online training courses via AWS Skill Builder are available to help developers get started, covering topics such as building conversational agents with Amazon Bedrock.

Sagemaker Jumpstart

Sagemaker Jumpstart has a lower learning curve as it emphasizes accessing pre-built models and solutions. Users can quickly deploy and evaluate models with a single click through the JumpStart UI or Python SDK. The platform also provides documentation, example notebooks, blogs, and video tutorials to assist users in applying various models and solutions to their specific use cases.

Amazon Q

The learning curve for Amazon Q varies based on intended usage. As an AI assistant, it aims to minimize entry barriers, allowing users to derive value with minimal training. For basic conversational assistance, users can begin interacting with Amazon Q immediately without any setup. It supports natural language communication via text.

For advanced capabilities, such as software development assistance, familiarity with tools like IDEs may be necessary to maximize its code-related features.

Integrating Amazon Q with organizational data and systems may present a steeper learning curve, involving the setup of connectors and plugins. Documentation and tutorials are available to facilitate this process.