Choosing a generative AI service

Amazon Q Business can answer questions, provide summaries, generate content, and securely complete tasks based on the data in your enterprise systems. It supports the general use case of using generative AI to start making the most of the information in your enterprise. With Amazon Q Business, you can make English-language queries about that information. It provides responses in a manner appropriate to your team’s needs. In addition, you can create lightweight, purpose-built Amazon Q Apps within your Amazon Q Business Pro subscription.

With Amazon Q Developer, you can understand, build, extend, and operate AWS applications. The supported use cases include tasks that range from coding, testing, and upgrading applications, to diagnosing errors, performing security scanning and fixes, and optimizing AWS resources. The advanced, multistep planning and reasoning capabilities in Amazon Q Developer are aimed at reducing the work involved in common tasks (such as performing Java version upgrades). These capabilities can also help implement new features generated from developer requests. 

Amazon Q Developer features are available as part of your workflows in other AWS services, including Amazon Q Developer in chat applications, Amazon CodeCatalyst, Amazon EC2, AWS Glue, and VPC Reachability Analyzer.

Chat with Amazon Q Developer to query and explore your AWS infrastructure directly from the AWS Management Console. Using natural language prompts to interact with your AWS account, you can get specific resource details and ask about relationships between resources.

Amazon Q in QuickSight is aimed at meeting the needs of a specific use case: getting actionable insights from your data by connecting Amazon Q to the Amazon Q QuickSight business intelligence (BI) service. You can use it to build visualizations of your data, summarize insights, answer data questions, and build data stories using natural language.

Amazon Q in Connect can automatically detect customer issues. It provides your customer service agents with contextual customer information along with suggested responses and actions for faster resolution of issues. It combines the capabilities of the Amazon Connect cloud contact center service with Amazon Q. Amazon Q in Connect can use your real-time conversations with your customers, along with relevant company content, to recommend what to say or what actions an agent should take to assist customers.

Model customization can deliver differentiated and personalized user experiences. To customize models for specific tasks, you can privately fine-tune FMs using your own labeled datasets. Custom models include capabilities such as fine-tuning and continued pre-training using unlabeled datasets. The list of FMs for which Amazon Bedrock supports fine-tuning includes Amazon Nova Micro, Lite, and Pro, Anthropic Claude 3 Haiku, Cohere Command, Meta Llama 2, Amazon Titan Text Lite and Express, Amazon Titan Multimodal Embeddings, and Amazon Titan Image Generator. The list of supported FMs is updated on an ongoing basis.

In addition, you can use Amazon Bedrock Custom Model Import to bring your own custom models and use them within Amazon Bedrock.

Amazon Bedrock Knowledge Bases is a fully managed capability that you can use to implement the entire Retrieval Augmented Generation (RAG) workflow—from ingestion to retrieval and prompt augmentation—without having to build custom integrations to data sources, and manage data flows. Session context management is built in, so your application can support multi-turn conversations. You can use the Retrieve API to fetch relevant results for a user query from knowledge bases.

With RAG, you can provide a model with new knowledge or up-to-date info from multiple sources, including document repositories, databases, and APIs. For example, the model might use RAG to retrieve search results from Amazon OpenSearch Service or documents from Amazon Simple Storage Service. Amazon Bedrock Knowledge Bases fully manages this experience by connecting to your private data sources, including Amazon Aurora, Amazon OpenSearch Serverless, MongoDB, Pinecone, and Redis Enterprise Cloud. This list includes connectors for Salesforce, Confluence, and SharePoint (in preview), so you can access more business data to customize models for your specific needs.

Amazon Bedrock Agents helps you plan and create multistep tasks using company systems and data sources—from answering customer questions about your product availability to taking their orders. You can create an agent by first selecting an FM and then providing it access to your enterprise systems, knowledge bases, and AWS Lambda functions to run your APIs securely. An agent analyzes the user request, and a Lambda function or your application can automatically call the necessary APIs and data sources to fulfill the request.

Agents can retain memory across multiple interactions to remember where you last left off and provide better recommendations based on prior interactions. Agents can also interpret code to tackle complex data-driven use cases, such as data analysis, data visualization, text processing, solving equations, and optimization problems.

Amazon Bedrock Guardrails evaluates user inputs and FM responses based on use case specific policies, and provides an additional layer of safeguards, regardless of the underlying FM. Using a short natural language description, you can use Amazon Bedrock Guardrails to define a set of topics to avoid within the context of your application. Guardrails detects and blocks user inputs and FM responses that fall into the restricted topics.

Guardrails supports contextual grounding checks, to detect hallucinations in model responses for applications using Retrieval Augmented Generation (RAG) and summarization applications. Contextual grounding checks add to the safety protection in Guardrails to make sure the LLM response is based on the right enterprise source data, and evaluates the LLM response to confirm that it’s relevant to the user’s query or instruction. Contextual grounding checks can detect and filter over 75% hallucinated responses for RAG and summarization workloads.

Use the Amazon Bedrock Converse API to create conversational applications that send and receive messages to and from an Amazon Bedrock model. For example, you can create a chatbot that maintains a conversation over many turns and uses a persona or tone customization that is unique to your needs, such as a helpful technical support assistant.

Tool use (function calling) gives a model access to tools that can help it generate responses for messages that you send to the model. For example, you might have a chat application that lets users find out the most popular song played on a radio station. To answer a request for the most popular song, a model needs a tool that can query and return the song information.

Use Amazon Bedrock IDE, a development experience in Amazon SageMaker Unified Studio (in preview), that lets you discover Amazon Bedrock models and build generative AI apps that use Amazon Bedrock models and features. For example, you can use a chat playground to try a prompt with an Anthropic Claude model without having to write any code. Later, you can use Amazon Bedrock IDE to create a generative AI app that uses an Amazon Bedrock model and features, such as a knowledge base or a guardrail, again without having to write any code.

Use Amazon Bedrock to create and save your own prompts using Prompt management, so that you can save time by applying the same prompt to different workflows. When you create a prompt, you can select a model to run inference on it and modify the inference parameters to use. You can include variables in the prompt so that you can adjust the prompt for different use case.

Prompt flows for Amazon Bedrock offers the ability for you to use supported FMs to build workflows by linking prompts, foundational models, and other AWS services to create comprehensive solutions.

With prompt flows, you can quickly build complex generative AI workflows using a visual builder. You can integrate with Amazon Bedrock offerings such as FMs, knowledge bases, and other AWS services such as AWS Lambda by transferring data between them. You can also deploy immutable workflows to move from testing to production in few clicks.

Amazon SageMaker JumpStart is an ML hub that provides access to publicly available foundation models. Those models include Mistral, Llama 3, CodeLlama, and Falcon 2. They can be customized with advanced fine-tuning and deployment techniques such as Parameter Efficient Fine-Tuning (PEFT) and Low-Rank Adaptation (LoRA).

This following screenshot shows some of the available models in Amazon SageMaker JumpStart within the AWS Management Console.

Amazon SageMaker Clarify addresses the all-important decision of which foundation model to use. Use Amazon SageMaker Clarify to create model evaluation jobs. A model evaluation job evaluates and compares model quality and responsibility metrics for text-based foundation models from JumpStart. Model evaluation jobs also support the use of JumpStart models that have already been deployed to an endpoint.

With Amazon SageMaker Canvas, you can use machine learning to generate predictions without writing any code. You can also use Amazon SageMaker Canvas in collaboration with Amazon Bedrock to fine-tune and deploy language models.

This blog post describes how you can use them to optimize customer interaction by working with your own datasets (such as your product FAQs) in Amazon Bedrock and Amazon SageMaker JumpStart The following diagram, from this blog post, demonstrates how Amazon SageMaker Canvas and Amazon Bedrock can be used together to fine-tune and deploy language models.

Amazon SageMaker Studio is a web-based experience for running ML workflows. Studio offers a suite of integrated development environments (IDEs). These include Code Editor, based on Code-OSS, Visual Studio Code - Open Source, a new JupyterLab application, RStudio, and Amazon SageMaker Studio Classic. For more information, see Applications supported in Amazon SageMaker Studio.

The web-based UI in Studio provides access to all SageMaker AI resources, including jobs and endpoints, in one interface. ML practitioners can also choose their preferred IDE to accelerate ML development. A data scientist can use JupyterLab to explore data and tune models. In addition, a machine learning operations (MLOps) engineer can use Code Editor with the pipelines tool in Studio to deploy and monitor models in production.

Amazon SageMaker Studio includes generative AI assistance, powered by Amazon Q Developer right within your JupyterLab Integrated Development Environment (IDE). With Q Developer, you can access expert guidance on SageMaker AI features, code generation, and troubleshooting.

Identify use cases/modality

What it is: Modality refers to the type of data the model processes: text, images (vision), or embeddings.

Why it matters: The choice of modality should align with the data that you're working with. For example, if your project involves processing natural language, a text-based model like Claude, Llama 3.1, or Titan Text G1 is suitable. If you want to create embeddings, then you might use a model like Titan Embeddings G1. Similarly, for image-related tasks, models such as Stable Diffusion XL, and Titan Image Generator v2, are more appropriate. Your use case might also involve considering your data source and the support for data source connectors, such as those provided in Amazon Q Business.

Model Size

What it is: This criterion refers to the number of parameters in a model. A parameter is a configuration variable that is internal to the model. Its values can be estimated (trained) during the training phase from the given training data. Parameters are crucial as they directly define the model's capability to learn from data. Large models often have more than 50 billion parameters.

Why it matters: The number of parameters is a key indicator of the model's complexity. More parameters mean that the model can capture more intricate patterns and nuances in the data, which generally leads to better performance. However, these models are not only expensive to train, but also require more computational resources to operate.

Inference latency

What it is: Inference speed, or latency, is the time it takes for a model to process input (often measured in tokens) and return an output. This processing time is crucial when the model's responses are part of an interactive system, like an Amazon Q Developer in chat applications.

Why it matters: Quick response times are essential for real-time applications such as interactive chatbots or instant translation services. These applications depend on the model's ability to process and respond to prompts rapidly to maintain a smooth user experience. Although larger FMs typically offer more detailed and accurate responses, their complex architectures can lead to slower inference speeds. This slower processing might frustrate users expecting immediate interaction.

To address this challenge, you can choose models optimized for quicker responses, even if it means compromising somewhat on the responses' depth or accuracy.

Maximizing context window

What it is: A large language model's context window is the amount of text (in tokens) that the model can consider at any one time when generating responses.

Why it matters: Larger context windows enable the model to remember and process more information in a single run. This ability is particularly valuable in complex tasks such as understanding long documents, engaging in detailed conversations, or generating contextually accurate text over larger spans.

For example, in a conversation, a model with a larger context window can remember more of the earlier dialogue, and provide responses that are more relevant to the entire conversation. This leads to a more natural and satisfying user experience, as the model can maintain the thread of discussion without losing context.

Pricing considerations

What it is: The cost of using an FM is influenced by the model's complexity and the model provider’s pricing structure.

Why it matters: Deploying high-performance models often comes with high costs due to increased computational needs. While these models provide advanced capabilities, their operational expenses can be high, particularly for startups or smaller projects on tight budgets.

Smaller, less resource-intensive models offer a more budget-friendly option without significantly compromising performance. Weigh the model’s cost against its benefits to ensure it fits within your project's financial constraints and gets you the best value for your investment.

Fine-tuning and continuous pre-training capability

What it is: Fine-tuning is a specialized training process in which a pre-trained model that has been trained on a large, generic dataset is further trained (or fine-tuned) on a smaller, specific dataset. This process adapts the model to particularities of the new data, improving its performance on related tasks. Continuous pre-training, on the other hand, involves extending the initial pre-training phase with additional training on new, emerging data that wasn't part of the original training set, helping the model stay relevant as data evolves. You can also use Retrieval Augmented Generation (RAG) to retrieve data from outside an FM and augment your prompts by adding the relevant retrieved data in context.

Why it matters: With fine-tuning, you can increase model accuracy by providing your own task-specific labeled training dataset and further specialize your FMs. With continued pre-training, you can train models using your own unlabeled data in a secure and managed environment. Continuous pre-training helps models become more domain-specific by accumulating more robust knowledge and adaptability beyond their original training.

Data quality

Data quality is a critical factor in the success of a generative AI application. Consider the following quality factors:

Data quantity

Quantity along with quality goes hand in hand. Consider the following quantity factors:

It's also important to continuously monitor and update your dataset as your generative AI applications evolve and as new data becomes available.

Quality of response

What it is: The most essential criterion is the quality of response. This is where you evaluate the output of a model based on several quality metrics, including accuracy, relevance, toxicity, fairness, and robustness against adversarial attacks.

Why it matters: The reliability and safety of model outputs are paramount, especially in applications that interact directly with users or make automated decisions that can affect people's lives. High-quality responses ensure user trust and satisfaction, reducing the risk of miscommunication and enhancing the overall user experience, thus earning the trust of your customers.