Monthly Archives: April 2025

Training a frontier model is highly compute-intensive, requiring a distributed system of hundreds, or thousands, of accelerated instances running for several weeks or months to complete a single job. For example, pre-training the Llama 3 70B model with 15 trillion training tokens took 6.5 million H100 GPU hours. On 256 Amazon EC2 P5 instances (p5.48xlarge, each with 8 NVIDIA H100 GPUs), this would take approximately 132 days.

Distributed training workloads run in a synchronous manner because each training step requires all participating instances to complete their calculations before the model can advance to the next step. It implies that if a single instance fails, it stops the entire job. As cluster sizes grow, the likelihood of failure increases due to the number of hardware components involved. Each hardware failure can result in wasted GPU hours and requires valuable engineering time to identify and resolve the issue, making the system prone to downtime that can disrupt progress and delay completion. To assess system reliability, engineering teams often rely on key metrics such as mean time between failures (MTBF), which measures the average operational time between hardware failures and serves as a valuable indicator of system robustness.

In this post, we explore the challenges of large-scale frontier model training, focusing on hardware failures and the benefits of Amazon SageMaker HyperPod—a resilient solution that minimizes disruptions, enhances efficiency, and reduces training costs.

Instance failure rate

To understand the typical MTBF for large-scale frontier model training, it helps to first understand instance failure rates by reviewing three noteworthy examples:

1. When training OPT-175B on 992 A100 GPUs, Meta AI encountered significant hardware reliability challenges. Across 2 months, the team managed 35 manual restarts and cycled over 100 hosts due to hardware issues, and automated systems triggered more than 70 restarts. Operating 124 instances (each with 8 GPUs) continuously over 1,440 hours, Meta accumulated a total of 178,560 instance-hours. The observed failure rate during this period was around 0.0588% per instance-hour, underscoring the reliability hurdles in training large frontier models at this scale.
2. During the training of Llama 3.1 405B on 16,000 H100 GPUs, a total of 417 unscheduled hardware failures occurred during a 54-day period. This translates to an effective failure rate of about 0.0161% per instance-hour.
3. MPT-7B was trained on 1 trillion tokens over the course of 9.5 days on 440 x A100-40GB. During this period, the training job experienced four hardware failures, resulting in an effective failure rate of approximately 0.0319% per instance-hour.

Based on these examples, it’s realistic to expect that in a single hour of large-scale distributed training, an instance will fail about 0.02%–0.06% of the time.

Larger clusters, more failures, smaller MTBF

As cluster size increases, the entropy of the system increases, resulting in a lower MTBF. The following table illustrates how the MTBF (in hours) changes with the number of instances in a cluster and the estimated failure rate for each instance. For example, with a 0.04% per-hour failure rate per instance, a 512-instance system is expected to experience a failure approximately every 5 hours. The following table shows MTBF (in hours) by failure rates.

Table 1: The change in MTBF (in hours) with the number of instances in a training cluster (with assumed failure rates in the columns)

What happens after a failure?

In a perfect world, without failures, the training job proceeds as shown in the following graph, which illustrates the total training time without failures, demonstrating a linear progression.

Figure 1: Training is linear in a perfect world without failures, since there are no interruptions to completion.

However, as previously noted, hardware failures are inevitable. Troubleshooting these failures typically involves several steps:

1. Root cause analysis (mean time to detect) – Identifying hardware failures as the root cause of training interruptions can be time-consuming, especially in complex systems with multiple potential failure points. The time taken to determine the root cause is referred to as mean time to detect (MTTD).
2. Hardware repair or replacement (mean time to replace) – Sometimes, a simple instance restart resolves the issue. At other times, the instance must be replaced, which can involve logistical delays, especially if specialized components aren’t readily available. If a replacement instance isn’t on hand when a GPU fails, the system must wait for one to become available. Common redistribution techniques, such as PyTorch FSDP, don’t permit workload redistribution among remaining instances.
3. System recovery and resumption (mean time to restart) – After resolving hardware issues and replacing the instance, additional time is needed to restore it to its previous state. The new instance must match the original configuration, and the entire cluster must load the model weights from the latest saved checkpoint.

Each failure incurs engineering effort to identify its root cause. When hardware issues arise, diagnostics confirm the problem and isolate the faulty instance, pausing the training job and increasing downtime. The impact of these failures is illustrated in the following figure and can be empirically measured for large distributed training jobs. The figure outlines the troubleshooting steps that follow a failure.

Figure 2: Impact of failures on a distributed training run. Once a failure occurs, time (idle GPUs) is spent on detecting (MTD), replacing (MTT Replace), and continuing (MTR Restart) a training run, often wasting time and expensive resources.

In a scenario where a distributed training job is running on an Amazon Elastic Kubernetes Service (Amazon EKS) cluster with n reserved instances and an Auto Scaling group set to maintain a minimum of n instances, a hardware issue such as a GPU failure can cause the job to fail. The affected instance will be marked as Unhealthy by a Kubernetes health monitor such as Node Problem Detector, and Amazon EKS will attempt to reschedule the training pods to healthy instances. If no instances have sufficient resources, the pods remain in a Pending state, and because the instance count is limited to n, no new instance will be automatically provisioned.

In such cases, the failed job must be manually identified through pod logs or the Kubernetes API and deleted. The failed instance also needs to be isolated and terminated manually, either through the AWS Management Console, AWS Command Line Interface (AWS CLI), or tools like kubectl or eksctl. To restore cluster capacity, the user must increase the cluster size by modifying the Auto Scaling group or updating the instance group. After the new instance is provisioned, bootstrapped, and added to the cluster, the training job must be restarted manually. If checkpointing is enabled, the job can resume from the last saved state. The overall downtime depends on the time required to provision a new instance and restart the job by rescheduling the pods.

Faster failure detection (shorter MTTD), shorter replacement times (shorter MTTR), and rapid resumption will all contribute to reducing total training time. Automating these processes with minimal user intervention is a key advantage of Amazon SageMaker HyperPod. 

Amazon SageMaker HyperPod resilient training infrastructure

SageMaker HyperPod is a compute environment optimized for large-scale frontier model training. This means users can build resilient clusters for machine learning (ML) workloads and develop or fine-tune state-of-the-art frontier models, as demonstrated by organizations such as Luma Labs and Perplexity AI. SageMaker HyperPod runs health monitoring agents in the background for each instance. When it detects a hardware failure, SageMaker HyperPod automatically repairs or replaces the faulty instance and resumes training from the last saved checkpoint. This automation alleviates the need for manual management, which means customers can train in distributed settings for weeks or months with minimal disruption. The benefits are particularly significant for customers deploying many instances (greater than 16) in a cluster.

Frontier model builders can further enhance model performance using built-in ML tools within SageMaker HyperPod. They can use Amazon SageMaker AI with MLflow to create, manage, and track ML experiments, or use Amazon SageMaker AI with TensorBoard to visualize model architecture and address convergence issues. Additionally, integrating with observability tools such as Amazon CloudWatch Container Insights, Amazon Managed Service for Prometheus, and Amazon Managed Grafana provides deeper insights into cluster performance, health, and utilization, ultimately saving valuable development time. The following figure compares the downtime of an infrastructure system using SageMaker HyperPod versus one without SageMaker HyperPod.

Figure 3: Comparing downtime chart from figure 1 with downtime on SageMaker HyperPod. When a failure occurs, it is detected automatically by HyperPod agents, and the instance is replaced in the background. Training is also resumed from the latest checkpoint

SageMaker HyperPod reduces the downtime per hardware failure by automatically detecting hardware issues. When these issues are detected, SageMaker HyperPod automatically replaces the faulty node(s) and resumes your training job from the latest checkpoint, assuming that checkpoints are written.

To evaluate this, we conducted experiments on SageMaker HyperPod using different cluster sizes of p5.48xlarge instances. The results in the following table, showing empirical measurements of time to resume by cluster size, displays the 90th percentile (P90), which represents a value that will be met or exceeded 90% of the time.

Table 2: MTTResume (in seconds) on clusters with different sizes

As shown, the mean time to replace an instance is independent of cluster size. For a cluster of 256 x p5.48xlarge instances training Meta Llama 3.1 70B parameter model with batch size = 8, replacing an instance takes about 940 seconds (or 15.7 minutes). After replacement, the new instance must install additional packages using lifecycle scripts and run deep health checks before reading from the latest saved checkpoint. When it’s operational, the training job resumes from the most recent checkpoint, minimizing progress loss despite the interruption. For a 256-instance cluster, it took us about 2,390 seconds (about 40 minutes) to automatically resume the training job after each failure.

Without SageMaker HyperPod, when a GPU failure occurs during a training job, the time it takes to resume the training can vary widely depending on the infrastructure and processes in place. With proper check-pointing, automated job orchestration, and efficient hardware provisioning, the resume time can be reduced. However, without these optimizations, the impact can be much more severe. Empirical evidence from customer experiences—including a leading open source frontier model provider, a top large language model (LLM) startup, an AI company specializing in enterprise frontier models, and a cutting-edge scientific research institute—indicates that without SageMaker HyperPod, the total downtime per GPU failure can average approximately 280 minutes per failure. Thus, Amazon SageMaker HyperPod saves about 240 minutes (or about 4 hours) of downtime per failure:

Table 3: Typical failure numbers, in minutes (as described in section “What happens after a failure?” with and without SageMaker HyperPod)

Quantifying the downtime savings

Depending on the frequency of failures, we can calculate the time to train and the cost savings of using SageMaker HyperPod. To illustrate this calculation, we assume it takes 40 minutes to replace an instance with SageMaker HyperPod compared to 280 minutes without it (as previously explained). Additionally, for this calculation, let’s assume a training job requiring 10 million GPU hours on H100 instances, running on a 256-instance P5 cluster.

Although the actual overhead (in hours) depends on the size of the training job, the relative overhead remains constant. The benefits of SageMaker HyperPod in reducing total training time are demonstrated in the following chart. For example, in a 256-instance cluster with a failure rate of 0.05%, SageMaker HyperPod reduces total training time by 32%.

Table 4: Total % of training time reduced by SageMaker HyperPod compared to a P5 cluster of comparable size

To translate this into actual savings, for a training job requiring 10 million GPU hours on a 256-instance cluster, SageMaker HyperPod saves 104 days of training time. As a result, customers can reduce time-to-market by 3.5 months. Without SageMaker HyperPod, the total time to train would be approximately 325 days, 121 of which are just spent on isolating and mitigating hardware issues. The following table shows the time to train benefits.

Table 5: Benefits presented by SageMaker HyperPod for a training run requiring 10 million GPU hours and a 256 instance cluster. SageMaker HyperPod saves 104 days of training time overall, resulting in a faster time to market (by 3.5 months!)

For the same example, we can estimate the total cost savings using:

Days lost due to hardware issues = (Number of instances) × (Failure rate per instance per hour) × (24 hours per day) × (Total training days) × (Downtime per failure in hours)

The following shows cost to train benefits.

Table 6: Using the calculation described above, the cost to train benefits laid out for a training run requiring 10 million GPU hours, 256 GPU based instances, and an assumed failure rate of 0.05% per instance per hour

A training job requiring 10 million GPU hours and 104 additional days of resolving hardware issues results in significant idle cluster time. Assuming a GPU cost of $5 per hour (equivalent to the price of P5 instances on Capacity Blocks for ML), the total cost savings with SageMaker HyperPod amounts to $25,559,040.

Summary

Training frontier models is a complex, resource-intensive process that is particularly vulnerable to hardware failures. In this post, we explored the instance failure rate, which can range about 0.02%–0.07% per hour during large-scale distributed training. As cluster size grows, the likelihood of failures increases, and the MTBF decreases. We also examined what happens after failure, including root cause analysis, hardware repair or replacement, and system recovery and resumption.

Next, we examined Amazon SageMaker HyperPod—a purpose-built, fully resilient cluster for frontier model training. By incorporating robust fault-tolerance mechanisms and automated health monitoring, SageMaker HyperPod minimizes disruptions caused by hardware issues. This not only streamlines the training process but also enhances the reliability and efficiency of model development, enabling faster and more effective innovation delivery. The benefits are measurable and correlate with both cluster size and failure rate. For a 256-instance cluster with a 0.05% per-instance-per-hour failure rate, SageMaker HyperPod reduces total training time by 32%, resulting in an approximate savings of $25.6 million in total training costs.

By addressing the reliability challenges of frontier model training, SageMaker HyperPod allows ML teams to focus on model innovation rather than infrastructure management. Organizations can now conduct long training runs with confidence, knowing that hardware failures will be automatically detected and resolved with minimal disruption to their ML workloads. Get started with Amazon SageMaker HyperPod.

Special thanks to Roy Allela, Senior AI/ML Specialist Solutions Architect for his support on the launch of this post.

About the Authors

Anoop Saha is a Sr GTM Specialist at Amazon Web Services (AWS) focusing on generative AI model training and inference. He partners with top frontier model builders, strategic customers, and AWS service teams to enable distributed training and inference at scale on AWS and lead joint GTM motions. Before AWS, Anoop held several leadership roles at startups and large corporations, primarily focusing on silicon and system architecture of AI infrastructure.

Trevor Harvey is a Principal Specialist in generative AI at Amazon Web Services (AWS) and an AWS Certified Solutions Architect – Professional. Trevor works with customers to design and implement machine learning solutions and leads go-to-market strategies for generative AI services.

Aman Shanbhag is a Specialist Solutions Architect on the ML Frameworks team at Amazon Web Services (AWS), where he helps customers and partners with deploying ML training and inference solutions at scale. Before joining AWS, Aman graduated from Rice University with degrees in computer science, mathematics, and entrepreneurship.

Read More

As businesses and developers increasingly seek to optimize their language models for specific tasks, the decision between model customization and Retrieval Augmented Generation (RAG) becomes critical. In this post, we seek to address this growing need by offering clear, actionable guidelines and best practices on when to use each approach, helping you make informed decisions that align with your unique requirements and objectives.

The introduction of Amazon Nova models represent a significant advancement in the field of AI, offering new opportunities for large language model (LLM) optimization. In this post, we demonstrate how to effectively perform model customization and RAG with Amazon Nova models as a baseline. We conducted a comprehensive comparison study between model customization and RAG using the latest Amazon Nova models, and share these valuable insights.

Approach and base model overview

In this section, we discuss the differences between a fine-tuning and RAG approach, present common use cases for each approach, and provide an overview of the base model used for experiments.

Demystifying RAG and model customization

RAG is a technique to enhance the capability of pre-trained models by allowing the model access to external domain-specific data sources. It combines two components: retrieval of external knowledge and generation of responses. It allows pre-trained language models to dynamically incorporate external data during the response-generation process, enabling more contextually accurate and updated outputs. Unlike fine-tuning, in RAG, the model doesn’t undergo any training and the model weights aren’t updated to learn the domain knowledge. Although fine-tuning implicitly uses domain-specific information by embedding the required knowledge directly into the model, RAG explicitly uses the domain-specific information through external retrieval.

Model customization refers to adapting a pre-trained language model to better fit specific tasks, domains, or datasets. Fine-tuning is one such technique, which helps in injecting task-specific or domain-specific knowledge for improving model performance. It adjusts the model’s parameters to better align with the nuances of the target task while using its general knowledge.

Common use cases for each approach

RAG is optimal for use cases requiring dynamic or frequently updated data (such as customer support FAQs and ecommerce catalogs), domain-specific insights (such as legal or medical Q&A), scalable solutions for broad applications (such as software as a service (SaaS) platforms), multimodal data retrieval (such as document summarization), and strict compliance with secure or sensitive data (such as financial and regulatory systems).

Conversely, fine-tuning thrives in scenarios demanding precise customization (such as personalized chatbots or creative writing), high accuracy for narrow tasks (such as code generation or specialized summarization), ultra-low latency (such as real-time customer interactions), stability with static datasets (such as domain-specific glossaries), and cost-efficient scaling for high-volume tasks (such as call center automation).

Although RAG excels at real-time grounding in external data and fine-tuning specializes in static, structured, and personalized workflows, choosing between them often depends on nuanced factors. This post offers a comprehensive comparison of RAG and fine-tuning, clarifying their strengths, limitations, and contexts where each approach delivers the best performance.

Introduction to Amazon Nova models

Amazon Nova is a new generation of foundation model (FM) offering frontier intelligence and industry-leading price-performance. Amazon Nova Pro and Amazon Nova Lite are multimodal models excelling in accuracy and speed, with Amazon Nova Lite optimized for low-cost, fast processing. Amazon Nova Micro focuses on text tasks with ultra-low latency. They offer fast inference, support agentic workflows with Amazon Bedrock Knowledge Bases and RAG, and allow fine-tuning for text and multi-modal data. Optimized for cost-effective performance, they are trained on data in over 200 languages.

Solution overview

To evaluate the effectiveness of RAG compared to model customization, we designed a comprehensive testing framework using a set of AWS-specific questions. Our study used Amazon Nova Micro and Amazon Nova Lite as baseline FMs and tested their performance across different configurations.

We structured our evaluation as follows:

• Base model:
  • Used out-of-box Amazon Nova Micro and Amazon Nova Lite
  • Generated responses to AWS-specific questions without additional context
• Base model with RAG:
  • Connected the base models to Amazon Bedrock Knowledge Bases
  • Provided access to relevant AWS documentation and blogs
• Model customization:
  • Fine-tuned both Amazon Nova models using 1,000 AWS-specific question-answer pairs generated from the same set of AWS articles
  • Deployed the customized models through provisioned throughput
  • Generated responses to AWS-specific questions with fine-tuned models

• Model customization and RAG combined approach:

• • Connected the fine-tuned models to Amazon Bedrock Knowledge Bases
  • Provided fine-tuned models access to relevant AWS articles at inference time

In the following sections, we walk through how to set up the second and third approaches (base model with RAG and model customization with fine-tuning) in Amazon Bedrock.

Prerequisites

To follow along with this post, you need the following prerequisites:

• An AWS account and appropriate permissions
• An Amazon Simple Storage Service (Amazon S3) bucket with two folders: one containing your training data, and one for your model output and training metrics

Implement RAG with the baseline Amazon Nova model

In this section, we walk through the steps to implement RAG with the baseline model. To do so, we create a knowledge base. Complete the following steps:

1. On the Amazon Bedrock console, choose Knowledge Bases in the navigation pane.
2. Under Knowledge Bases, choose Create.

3. On the Configure data source page, provide the following information:
   1. Specify the Amazon S3 location of the documents.
   2. Specify a chunking strategy.
4. Choose Next.

5. On the Select embeddings model and configure vector store page, provide the following information:
   1. In the Embeddings model section, choose your embeddings model, which is used for embedding the chunks.
   2. In the Vector database section, create a new vector store or use an existing one where the embeddings will be stored for retrieval.
6. Choose Next.

7. On the Review and create page, review the settings and choose Create Knowledge Base.

Fine-tune an Amazon Nova model using the Amazon Bedrock API

In this section, we provide detailed walkthroughs on fine-tuning and hosting customized Amazon Nova models using Amazon Bedrock. The following diagram illustrates the solution architecture.

Create a fine-tuning job

Fine-tuning Amazon Nova models through the Amazon Bedrock API is a streamlined process:

1. On the Amazon Bedrock console, choose us-east-1 as your AWS Region.

At the time of writing, Amazon Nova model fine-tuning is exclusively available in us-east-1.

2. Choose Custom models under Foundation models in the navigation pane.
3. Under Customization methods, choose Create Fine-tuning job.

4. For Source model, choose Select model.
5. Choose Amazon as the provider and the Amazon Nova model of your choice.
6. Choose Apply.

7. For Fine-tuned model name, enter a unique name for the fine-tuned model.
8. For Job name, enter a name for the fine-tuning job.
9. Under Input data, enter the location of the source S3 bucket (training data) and target S3 bucket (model outputs and training metrics), and optionally the location of your validation dataset.

Configure hyperparameters

For Amazon Nova models, the following hyperparameters can be customized:

Prepare the dataset for compatibility with Amazon Nova models

Similar to other LLMs, Amazon Nova requires prompt-completion pairs, also known as question and answer (Q&A) pairs, for supervised fine-tuning (SFT). This dataset should contain the ideal outputs you want the language model to produce for specific tasks or prompts. Refer to Guidelines for preparing your data for Amazon Nova on best practices and example formats when preparing datasets for fine-tuning Amazon Nova models.

Examine fine-tuning job status and training artifacts

After you create your fine-tuning job, choose Custom models under Foundation models in the navigation pane. You will find the current fine-tuning job listed under Jobs. You can use this page to monitor your fine-tuning job status.

When your fine-tuning job status changes to Complete, you can choose the job name and navigate to the Training job overview page. You will find the following information:

• Training job specifications
• Amazon S3 location for input data used for fine-tuning
• Hyperparameters used during fine-tuning
• Amazon S3 location for training output

Host the fine-tuned model with provisioned throughput

After your fine-tuning job completes successfully, you can access your customized model through the following steps:

1. On the Amazon Bedrock console, choose Custom models under Foundation models in the navigation pane.
2. Under Models, choose your custom model.

The model details page shows the following information:

• Fine-tuned model details
• Amazon S3 location for input data used for fine-tuning
• Hyperparameters used during fine-tuning
• Amazon S3 location for training output

3. To make your fine-tuned model available for inference, choose Purchase provisioned throughput.
4. Choose a commitment term (no commitment, 1 month, or 6 months) and review the associated cost for hosting the fine-tuned models.

After the customized model is hosted through provisioned throughput, a model ID will be assigned and can be used for inference.

The aforementioned fine-tuning and inference steps can also be done programmatically. For more information, refer to the following GitHub repo, which contains sample code.

Evaluation framework and results

In this section, we first introduce our multi-LLM-judge evaluation framework, which is set up to mitigate an individual LLM judge’s bias. We then compare RAG vs. fine-tuning results in terms of response quality as well as latency and token implications.

Multiple LLMs as judges to mitigate bias

The following diagram illustrates our workflow using multiple LLMs as judges.

Using LLMs as judges has become an increasingly popular approach to evaluate tasks that are challenging to assess through traditional methods or human evaluation. For our evaluation framework, we constructed 10 domain-specific test questions covering key aspects of AWS services and features, designed to test both factual accuracy and depth of understanding. Each model-generated response was evaluated using a standardized scoring system on a scale of 0–10, where 0–3 indicates incorrect or misleading information, 4–6 represents partially correct but incomplete answers, 7–8 signifies mostly correct with minor inaccuracies, and 9–10 denotes completely accurate with comprehensive explanation.

We use the following LLM judge evaluation prompt:

{
    "system_prompt": "You are a helpful assistant.",
    "prompt_template": "[Instruction] Please act as an impartial judge and evaluate the quality of the response provided by an AI assistant to the user question displayed below. Your evaluation should consider factors such as the helpfulness, relevance, accuracy, depth, creativity, and level of detail of the response. Begin your evaluation by providing a short explanation. Be as objective as possible. After providing your explanation, you must rate the response on a scale of 1 to 10 by strictly following this format: "[[rating]]", for example: "Rating: [[5]]".nn[Question]n{question}nn[The Start of Assistant's Answer]n{answer}n[The End of Assistant's Answer]",
    "description": "Prompt for general questions",
    "category": "general",
    "output_format": "[[rating]]"
}

We use the following sample evaluation question and ground truth:

{
    "question_id": 9161,
    "category": "AWS",
    "turns": [
        " "What specific details are collected and sent to AWS when anonymous operational metrics are enabled for an Amazon EFS file system?",
        "What's required for a successful AWS CloudFormation launch?"
    ],
    "reference": [
        "When anonymous operational metrics are enabled for an Amazon EFS file system, the following specific details are collected and sent to AWS: Solution ID, Unique ID, Timestamp, Backup ID, Backup Start Time, Backup Stop Time, Backup Window, Source EFS Size, Destination EFS Size, Instance Type, Retain, S3 Bucket Size, Source Burst Credit Balance, Source Burst Credit Balance Post Backup, Source Performance Mode, Destination Performance Mode, Number of Files, Number of Files Transferred, Total File Size, Total Transferred File Size, Region, Create Hard Links Start Time, Create Hard Links Stop Time, Remove Snapshot Start Time, Remove Snapshot Stop Time, Rsync Delete Start Time, Rsync Delete Stop Time.",
        "For a successful AWS CloudFormation launch, you need to sign in to the AWS Management Console, choose the correct AWS Region, use the button to launch the template, verify the correct template URL, assign a name to your solution stack, review and modify the parameters as necessary, review and confirm the settings, check the boxes acknowledging that the template creates AWS Identity and Access Management resources and may require an AWS CloudFormation capability, and choose Create stack to deploy the stack. You should receive a CREATE_COMPLETE status in approximately 15 minutes."
    ]
}

To mitigate potential intrinsic biases among different LLM judges, we adopted two LLM judges to evaluate the model-generated responses: Anthropic’s Claude Sonnet 3.5 and Meta’s Llama 3.1 70B. Each judge was provided with the original test question, the model-generated response, and specific scoring criteria focusing on factual accuracy, completeness, relevance, and clarity. Overall, we observed a high level of rank correlation among LLM judges in assessing different approaches, with consistent evaluation patterns across all test cases.

Response quality comparison

Both fine-tuning and RAG significantly improve the quality of generated responses on AWS-specific questions over the base model. Using Amazon Nova Lite as the base model, we observed that both fine-tuning and RAG improved the average LLM judge score on response quality by 30%, whereas combining fine-tuning with RAG enhanced the response quality by a total of 83%, as shown in the following figure.

Notably, our evaluation revealed an interesting finding (as shown in the following figure): when combining fine-tuning and RAG approaches, smaller models like Amazon Nova Micro showed significant performance improvements in domain-specific tasks, nearly matching the performance of bigger models. This suggests that for specialized use cases with well-defined scope, using smaller models with both fine-tuning and RAG could be a more cost-effective solution compared to deploying larger models.

Latency and token implications

In addition to enhancing the response quality, both fine-tuning and RAG help reduce the response generation latency compared to the base model. For both Amazon Nova Micro and Amazon Nova Lite, fine-tuning reduced the base model latency by approximately 50%, whereas RAG reduced it by about 30%, as shown in the following figure.

Fine-tuning also presented the unique advantage of improving the tone and style of the generated answers to align more closely with the training data. In our experiments, the average total tokens (input and output tokens) dropped by more than 60% with both fine-tuned models. However, the average total tokens more than doubled with the RAG approach due to passing of context, as shown in the following figure. This finding suggests that for latency-sensitive use cases or when the objective is to align the model’s responses to a specific tone, style, or brand voice, model customization might offer more business value.

Conclusion

In this post, we compared model customization (fine-tuning) and RAG for domain-specific tasks with Amazon Nova. We first provided a detailed walkthrough on how to fine-tune, host, and conduct inference with customized Amazon Nova through the Amazon Bedrock API. We then adopted an LLM-as-a-judge approach to evaluate response quality from different approaches. In addition, we examined the latency and token implications of different setups.

Both fine-tuning and RAG improved the model performance. Depending on the task and evaluation criteria, model customization showed similar, or sometimes better, performance compared to RAG. Model customization can also be helpful to improve the style and tone of a generated answer. In this experiment, the customized model’s response follows the succinct answer style of the given training data, which resulted in lower latency compared to the baseline counterpart. Additionally, model customization can also be used for many use cases where RAG isn’t as straightforward to be used, such as tool calling, sentiment analysis, entity extraction, and more. Overall, we recommend combining model customization and RAG for question answering or similar tasks to maximize performance.

For more information on Amazon Bedrock and the latest Amazon Nova models, refer to the Amazon Bedrock User Guide and Amazon Nova User Guide. The AWS Generative AI Innovation Center has a group of AWS science and strategy experts with comprehensive expertise spanning the generative AI journey, helping customers prioritize use cases, build a roadmap, and move solutions into production. Check out the Generative AI Innovation Center for our latest work and customer success stories.

About the Authors

Mengdie (Flora) Wang is a Data Scientist at AWS Generative AI Innovation Center, where she works with customers to architect and implement scalable Generative AI solutions that address their unique business challenges. She specializes in model customization techniques and agent-based AI systems, helping organizations harness the full potential of generative AI technology. Prior to AWS, Flora earned her Master’s degree in Computer Science from the University of Minnesota, where she developed her expertise in machine learning and artificial intelligence.

Sungmin Hong is a Senior Applied Scientist at Amazon Generative AI Innovation Center where he helps expedite the variety of use cases of AWS customers. Before joining Amazon, Sungmin was a postdoctoral research fellow at Harvard Medical School. He holds Ph.D. in Computer Science from New York University. Outside of work, he prides himself on keeping his indoor plants alive for 3+ years.

Jae Oh Woo is a Senior Applied Scientist at the AWS Generative AI Innovation Center, where he specializes in developing custom solutions and model customization for a diverse range of use cases. He has a strong passion for interdisciplinary research that connects theoretical foundations with practical applications in the rapidly evolving field of generative AI. Prior to joining Amazon, Jae Oh was a Simons Postdoctoral Fellow at the University of Texas at Austin, where he conducted research across the Mathematics and Electrical and Computer Engineering departments. He holds a Ph.D. in Applied Mathematics from Yale University.

Rahul Ghosh is an Applied Scientist at Amazon’s Generative AI Innovation Center, where he works with AWS customers across different verticals to expedite their use of Generative AI. Rahul holds a Ph.D. in Computer Science from the University of Minnesota.

Baishali Chaudhury is an Applied Scientist at the Generative AI Innovation Center at AWS,
where she focuses on advancing Generative AI solutions for real-world applications. She has a
strong background in computer vision, machine learning, and AI for healthcare. Baishali holds a PhD in Computer Science from University of South Florida and PostDoc from Moffitt Cancer Centre.

Anila Joshi has more than a decade of experience building AI solutions. As a AWSI Geo Leader at AWS Generative AI Innovation Center, Anila pioneers innovative applications of AI that push the boundaries of possibility and accelerate the adoption of AWS services with customers by helping customers ideate, identify, and implement secure generative AI solutions.

Read More

Today, businesses are using AI and generative models to improve productivity in their teams and provide better experiences to their customers. Personalized outbound communication can be a powerful tool to increase user engagement and conversion.

For instance, as a marketing manager for a video-on-demand company, you might want to send personalized email messages tailored to each individual user—taking into account their demographic information, such as gender and age, and their viewing preferences. You want the messaging and movie recommendations to be both engaging and applicable to the customer. To achieve this, you can use Amazon Personalize to generate user-personalized recommendations and Amazon Bedrock to generate the text of the email.

Amazon Personalize enables your business to improve customer engagement by creating personalized product and content recommendations in websites, applications, and targeted marketing campaigns. You can get started without any prior machine learning (ML) experience, and Amazon Personalize allows you to use APIs to build sophisticated personalization capabilities. Using this service, all your data is encrypted to be private and secure, and is only used to create recommendations for your users.

Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models (FMs) from leading AI companies and Amazon through a single API, along with a broad set of capabilities to build generative AI applications with security, privacy, and responsible AI. Using Amazon Bedrock, you can experiment with and evaluate top FMs for your use case, customize the model using fine tuning, or restrict the model output using Retrieval Augmented Generaion (RAG), and build agents that execute tasks using your enterprise systems and data sources.

In this post, we demonstrate how to use Amazon Personalize and Amazon Bedrock to generate personalized outreach emails for individual users using a video-on-demand use case. This concept can be applied to other domains, such as compelling customer experiences for ecommerce and digital marketing use cases.

Solution overview

The following diagram shows how you can use Amazon Personalize and Amazon Bedrock to generate user-personalized outreach messages for each user.

The workflow consists of the following steps:

1. Import your user, item, and interaction data into Amazon Personalize. The user and item datasets are not required for Amazon Personalize to generate recommendations, but providing good item and user metadata provides the best results in your trained models.
2. Train an Amazon Personalize “Top picks for you” recommender. Amazon Personalize recommenders are domain-specific resources that generate recommendations. When you create an Amazon Personalize recommender, Amazon Personalize trains the models backing the recommender with the best configurations for the use case. In our example, we use the “Top picks for you” recommender. This recommender generates personalized content recommendations for a user that you specify. With this use case, Amazon Personalize automatically filters videos the user watched.
3. After the model is trained, you can get the top recommended movies for each user by querying the recommender with each user ID through the Amazon Personalize Runtime API.
4. Combine a predefined prompt template with the top recommendations and user demographic information to generate an enhanced prompt.
5. Use the enhanced prompt in Amazon Bedrock through its API to generate your personalized outbound communication.
6. Amazon Bedrock returns the personalized outbound communication that you can email to your users.

We go deeper into each of these steps in the following sections. A code sample for this use case is available on AWS Samples on GitHub.

Prerequisites

To generate personalized recommendations, you must first set up Amazon Personalize resources. You start by creating your dataset group, loading your data, and then training a recommender. For full instructions, see Getting started tutorials.

1. 1. Create a dataset group.
   2. Create an Interactions dataset using the following schema:

      {
          "type": "record"
          "name": "Interactions",
          "namespace": "com.amazonaws.personalize.schema",
          "fields": [
              {
                  "name": "USER_ID",
                  "type": "string"
              },
              {
                  "name": "ITEM_ID",
                  "type": "string"
              },
              {
                  "name": "TIMESTAMP",
                  "type": "long"
              },
              {
                  "name": "EVENT_TYPE",
                  "type": "string"
              }
          ],
          "version": "1.0"
      }

      Interaction data consists of information about the user interactions with the content in your application. This usually comes from analytics tools or a customer data platform (CDP). The best interaction data to use in Amazon Personalize includes the sequential order of user behavior and the content the user watched or clicked on. For this example, we use the ml-latest-small dataset from the MovieLens dataset to simulate user-item interactions.

   3. Import the interaction data to Amazon Personalize from Amazon Simple Storage Service (Amazon S3). For this example, we convert the data to the appropriate format following the steps in the notebook 01_Introduction_and_Data_Preparation.
   4. Item data consists of information about the content that is being interacted with, which generally comes from a content management system (CMS) in video-on-demand use cases. This can be information like the title, description, or movie genre. To provide additional metadata, and also provide a consistent experience for our users, we use a subset of the IMDb Essential Metadata for Movies/TV/OTT dataset. IMDb has multiple datasets available in AWS Data Exchange. For this post, we have extracted and prepared a subset of data for use with the following information from the IMDb Essential Metadata for Movies/TV/OTT (Bulk data) dataset.With this data, create an Items dataset using the following schema:

      items_schema = {
          "type": "record",
          "name": "Items",
          "namespace": "com.amazonaws.personalize.schema",
          "fields": [
              {
                  "name": "ITEM_ID",
                  "type": "string"
              },
              {
                  "name": "TITLE",
                  "type": "string"
              },
              {
                  "name": "YEAR",
                  "type": "int"
              },
              {
                  "name": "IMDB_RATING",
                  "type": "int"
              },
              {
                  "name": "IMDB_NUMBEROFVOTES",
                  "type": "int"
              },
              {
                  "name": "PLOT",
                  "type": "string",
                  "textual": True
              },
              {
                  "name": "US_MATURITY_RATING_STRING",
                  "type": "string"
              },
              {
                  "name": "US_MATURITY_RATING",
                  "type": "int"
              },
              {
                  "name": "GENRES",
                  "type": "string",
                  "categorical": True
              },
              {
                  "name": "CREATION_TIMESTAMP",
                  "type": "long"
              },
              {
                  "name": "PROMOTION",
                  "type": "string"
              }
          ],
          "version": "1.0
      }

   5. Import the item data to Amazon Personalize from Amazon S3. For this example, we convert the data to the appropriate format following the steps in the notebook 01_Introduction_and_Data_Preparation.
      For more information on formatting and importing your interactions and items data from Amazon S3, see Importing bulk records.
   6. Create a recommender. In this example, we create a “Top picks for you” recommender.

Get personalized recommendations using Amazon Personalize

Now that we have trained the “Top picks for you” recommender, we can generate recommendations for our users. For more details and ways to use Amazon Personalize to get recommendations, see Getting recommendations from Amazon Personalize.We include the item metadata in the response so we can use this information in our outbound communication in the next step.You can use the following code to get recommended movies for each user:

get_recommendations_response = personalize_runtime.get_recommendations(
    recommenderArn = workshop_recommender_top_picks_arn,
    userId = str(user_id),
    numResults = number_of_movies_to_recommend,
    metadataColumns = {
        "ITEMS": [
            'TITLE', 'PLOT', 'GENRES']
        }
)

In the items dataset, we can specify the metadata columns we want the recommender to return. In this case, we request the Title, Plot, and Genres of the recommended movie. You can request metadata columns only if this feature has been enabled when the recommender was created.

For an example user_Id, the following movies are recommended:

Title: There's Something About Mary
Genres: Comedy and Romance
Plot: A man gets a chance to meet up with his dream girl from high school, even though his date with her back then was a complete disaster.

Title: Shakespeare in Love
Genres: Comedy and Drama and History and Romance
Plot: The world's greatest ever playwright, William Shakespeare, is young, out of ideas and short of cash, but meets his ideal woman and is inspired to write one of his most famous plays.

Title: The Birdcage
Genres: Comedy
Plot: A gay cabaret owner and his drag queen companion agree to put up a false straight front so that their son can introduce them to his fiancée's right-wing moralistic parents.

Get the user’s favorite movie genre

To provide a better personalized outbound communication experience, we determine the user’s favorite movie genre based on the genres of all the movies they have interacted with in the past. There are a number of ways to do this, such as counting the number of interactions per genre for our user. In this example, our sample user’s favorite genre is Comedy.

Generate personalized marketing emails with recommended movies

To generate personalized marketing emails, we use Amazon Bedrock. Amazon Bedrock users must request access to models before they are available for use. Amazon Bedrock is a fully managed service that makes base models from Amazon and third-party model providers accessible through an API.

To request access, select choose Model access in the navigation pane on the Amazon Bedrock console. For more information, see Access Amazon Bedrock foundation models.

In this example, we use Anthropic’s Claude 3.7 on Amazon Bedrock and have defined the following configuration parameters:

# The LLM we will be using
model_id = 'us.anthropic.claude-3-7-sonnet-20250219-v1:0'

# The maximum number of tokens to use in the generated response
max_tokens_to_sample = 1000

Let’s generate a simple outreach email using the recommended movies and the following prompt template:

prompt_template = f'''Write a marketing email advertising several movies available in a video-on-demand streaming platform next week, given the movie and user information below. The movies to recommend and their information is contained in the <movie> tag. Put the email between <email> tags.

<movie>
{movie_list}
</movie>

Assistant: Email body:
<email>
'''

Using the recommended movies, the full prompt is as follows:

"Write a marketing email advertising several movies available in a video-on-demand streaming platform next week, given the movie and user information below. The movies to recommend and their information is contained in the <movie> tag. Put the email between <email> tags.
n
n
<movie>
n
[
{
'title': "There's Something About Mary",
'genres': 'Comedy and Romance',
'plot': 'A man gets a chance to meet up with his dream girl from high school, even though his date with her back then was a complete disaster.'
},
{
'title': 'Shakespeare in Love',
'genres': 'Comedy and Drama and History and Romance',
'plot': "The world's greatest ever playwright, William Shakespeare, is young, out of ideas and short of cash, but meets his ideal woman and is inspired to write one of his most famous plays."
},
{
'title': 'The Birdcage',
'genres': 'Comedy',
'plot': "A gay cabaret owner and his drag queen companion agree to put up a false straight front so that their son can introduce them to his fiancu00e9e's right-wing moralistic parents."
}
]
n
</movie>
n
n
Assistant: Email body:
n
<email>.
"

We then use an Amazon Bedrock API call to generate the personalized email. For more information, see Amazon Bedrock API Reference.

request_body = json.dumps({
    "max_tokens": max_tokens_to_sample,
    "messages": [{"role": "user", "content": prompt}],
    "anthropic_version": "bedrock-2023-05-31"
})

personalized_email_response = bedrock_client.invoke_model(
    body = request_body,
    modelId = identifier_of_the_model
)

Amazon Bedrock returns a personalized email for the user:

Subject: Your Weekend Movie Escape Awaits! Three Laugh-Out-Loud Comedies Coming Next Week

Hi there,

Need a break from reality? We’ve got you covered with three fantastic comedies hitting our streaming platform next week!

## This Week’s Spotlight: Comedy Gems That Will Make Your Day

**There’s Something About Mary**
This hilarious romantic comedy follows a man who finally gets a second chance with his high school dream girl—after their first date went hilariously wrong. With unforgettable laughs and heartwarming moments, it’s the perfect weekend watch!

**Shakespeare in Love**
When the greatest playwright of all time faces writer’s block and money troubles, an unexpected romance changes everything! This award-winning comedy-drama blends history, romance, and witty humor as Shakespeare finds his muse and creates one of his most beloved plays. A delightful escape for literature lovers and romantics alike!

**The Birdcage**
Prepare for non-stop laughter in this comedy classic! When a gay cabaret owner and his drag queen partner pretend to be straight to impress their future in-laws (who happen to be ultra-conservative), chaos and hilarity ensue. A perfect blend of humor and heart that still resonates today.

So grab your popcorn, get comfortable on the couch, and enjoy these comedy classics starting next week!

Happy streaming!

The Movies-On-Demand Team

P.S. Don’t forget to check out our complete catalog for more great films in every genre!

Although this is already a good outreach email because the recommendations are personalized to the user, we can personalize it further by adding more information about the user.

Generate personalized communication with recommended movies, user demographic information, and favorite genre

We will generate emails by assuming two different demographics for the users as well as their favorite genre.

The version of the ml-latest-small dataset from the MovieLens dataset we used in this example doesn’t contain demographic data; therefore, we will try out multiple options. In a real-world scenario, you might know the demographics of your audience.

To experiment, let’s use the following example demographic:

# Sample user demographics
user_demographic_1 = f'The user is a 50 year old adult called Otto.'

We also add the user’s favorite genre to the prompt as follows:

prompt_template = f'''You are a skilled publicist. Write a high-converting marketing email advertising several movies available in a video-on-demand streaming platform next week,
given the movie and user information below. Do not add additional information. Your email will leverage the power of storytelling and persuasive language.
You want the email to impress the user, so make it appealing to them based on the information contained in the <user> tags,
and take into account the user's favorite genre in the <genre> tags.
The movies to recommend and their information is contained in the <movie> tag.
All movies in the <movie> tag must be recommended. Give a summary of the movies and why the human should watch them.
Put the email between <email> tags.

<user>
{user_demographic}
</user>

<genre>
{favorite_genre}
</genre>

<movie>
{movie_list}
</movie>

Assistant:

<email>
'''

After adding the information, the new prompt is as follows:

"You are a skilled publicist. Write a high-converting marketing email advertising several movies available in a video-on-demand streaming platform next week, given the movie and user information below. Do not add additional information. Your email will leverage the power of storytelling and persuasive language. You want the email to impress the user, so make it appealing to them based on the information contained in the <user> tags, and take into account the user's favorite genre in the <genre> tags. The movies to recommend and their information is contained in the <movie> tag. All movies in the <movie> tag must be recommended. Give a summary of the movies and why the human should watch them. Put the email between <email> tags.
n
n
<user>
n
The user is a 50 year old adult called Otto.
n
</user>
n
n
<genre>
n
Comedy
n
</genre>
n
n
<movie>
n
[
{
'title': "There's Something About Mary",
'genres': 'Comedy and Romance',
'plot': 'A man gets a chance to meet up with his dream girl from high school, even though his date with her back then was a complete disaster.'
},
{
'title': 'Shakespeare in Love',
'genres': 'Comedy and Drama and History and Romance',
'plot': "The world's greatest ever playwright, William Shakespeare, is young, out of ideas and short of cash, but meets his ideal woman and is inspired to write one of his most famous plays."
},
{
'title': 'The Birdcage',
'genres': 'Comedy',
'plot': "A gay cabaret owner and his drag queen companion agree to put up a false straight front so that their son can introduce them to his fiancu00e9e's right-wing moralistic parents."
}
]
n
</movie>
n
n
Assistant:
n
<email>
n    "

Amazon Bedrock returns a personalized email for the user:

Subject: Otto, Get Ready for a Comedy Extravaganza on Your Screen Next Week!

Dear Otto,

We’re thrilled to bring you an exclusive lineup of comedy classics hitting our streaming platform next week! As someone who appreciates a good laugh, you’re in for a treat with these award-winning comedies that will brighten your evenings.

## “There’s Something About Mary”
This hilarious romantic comedy follows the misadventures of a man who finally gets a second chance with his high school dream girl. After a first date that was nothing short of catastrophic, he’s determined to make things right years later. With its perfect blend of outrageous humor and heartwarming moments, this comedy classic delivers laughs that have stood the test of time.

## “Shakespeare in Love”
Experience the witty and charming story of a young, broke William Shakespeare who finds his muse in the most unexpected place. This brilliant comedy-drama offers a fictional account of how the greatest playwright found inspiration through love. With its clever dialogue, historical setting, and romantic storyline, this Academy Award-winning film combines your love of comedy with rich storytelling that will keep you engaged from beginning to end.

## “The Birdcage”
A comedy masterpiece that delivers non-stop laughs! When a gay cabaret owner and his flamboyant partner must pretend to be straight to impress their future in-laws (who happen to be ultra-conservative), chaos ensues. The brilliant performances and hilarious situations make this one of the most beloved comedies of its era. It’s the perfect film for when you need genuine belly laughs and brilliant comedic timing.

Otto, these comedies are among the best in their genre and will be available for your enjoyment starting next week. Whether you’re in the mood for slapstick humor, clever wit, or situational comedy, this collection has something perfect for your evening entertainment.

Grab your favorite snack, get comfortable on the couch, and prepare for an unforgettable comedy marathon!

Happy streaming!

The VOD Team

The email now contains information about the user’s favorite genre and is personalized to the user using their name and the recommended the movies the user is most likely to be interested in.

Clean up

Make sure you clean up any unused resources you created in your account while following the steps outlined in this post. You can delete filters, recommenders, datasets, and dataset groups using the AWS Management Console or the Python SDK.

Conclusion

Traditional AI and generative AI allow you to build hyper-personalized experiences for your users. In this post, we showed how to generate personalized outbound communication by getting personalized recommendations for each user using Amazon Personalize and then using user preferences and demographic information to write a personalized email communication using Amazon Bedrock. By using AWS managed services, such as Amazon Personalize and Amazon Bedrock, you can create this content with only a few API calls—no ML experience required.

For more information about Amazon Personalize, see the Amazon Personalize Developer Guide. For more information on working with generative AI on AWS, see Announcing New Tools for Building with Generative AI on AWS.

About the Author

Anna Grüebler Clark is a Specialist Solutions Architect at AWS focusing on in Artificial Intelligence. She has more than 16 years experience helping customers develop and deploy machine learning applications. Her passion is taking new technologies and putting them in the hands of everyone, and solving difficult problems leveraging the advantages of using traditional and generative AI in the cloud.

Read More

Financial institutions today face an increasingly complex regulatory world that demands robust, efficient compliance mechanisms. Although organizations traditionally invest countless hours reviewing regulations such as the Anti-Money Laundering (AML) rules and the Bank Secrecy Act (BSA), modern AI solutions offer a transformative approach to this challenge. By using Amazon Bedrock Knowledge Bases alongside CrewAI—an open source multi-agent orchestration framework, organizations can now deploy intelligent systems where multiple AI agents work together to automate and streamline specific compliance processes. This powerful combination enables financial institutions to move from manual, time-intensive compliance reviews to a streamlined, assisted compliance management approach that adapts to evolving regulatory requirements.

In this post, we explore how AI agents can streamline compliance and fulfill regulatory requirements for financial institutions using Amazon Bedrock and CrewAI. We demonstrate how to build a multi-agent system that can automatically summarize new regulations, assess their impact on operations, and provide prescriptive technical guidance. You’ll learn how to use Amazon Bedrock Knowledge Bases and Amazon Bedrock Agents with CrewAI to create a comprehensive, automated compliance solution.

This solution’s architecture can be adapted to help healthcare systems, enable manufacturers to maintain ISO safety documentation, and assist retailers in monitoring Federal Trade Commission (FTC) advertising regulations. It can also assist in other segments such as legal, finance, or human resources, offering wide-ranging potential for process automation and efficiency gains across various industries.The code used for this post is available on GitHub.

Solution overview

Traditional large language model (LLM) applications excel at following predefined instructions, but solving complex challenges such as compliance automation requires an autonomous network of specialized agents that mirror the structure of a comprehensive compliance department. Our system employs three key agents:

1. Compliance analyst agent that continuously monitors and analyzes regulatory changes
2. Compliance specialist agent that transforms requirements into organizational policies
3. Enterprise architect agent that designs and implements the necessary security controls

In this multi-agent approach, specialized AI agents work together seamlessly to streamline the compliance lifecycle. The compliance analyst agent collects latest regulatory changes and helps to stay ahead of regulatory changes and their potential impact where the Compliance specialist agent translates these regulatory requirements into actionable organizational procedures. Meanwhile, the enterprise architect agent makes sure that the technical controls align with organizational controls. CrewAI provides an open source framework to orchestrate this collaborative system, enabling these agents to work in concert while maintaining clear handoffs and accountability. Next, we will explore how to create this multi-agent compliance automation system using CrewAI.

Although this solution demonstrates CrewAI’s capabilities, it’s important to note that Amazon Bedrock Agents has built-in support for multi-agent collaboration, and organizations could implement their agent workflows entirely within Amazon Bedrock Agents. However, we’ve chosen CrewAI for this demonstration to showcase how open source frameworks can extend Amazon Bedrock capabilities while maintaining enterprise-grade security through Bedrock Guardrails.

Solution components

This solution shows you how to combine multiple capabilities. It shows how to:

1. Develop a multi-agent solution using a CrewAI framework
2. Enrich the solution using domain-specific data using Amazon Bedrock Knowledge Bases
3. Safeguard your generative AI application using Amazon Bedrock Guardrails
4. Bring everything together using CrewAI and Amazon Bedrock Agents

You can use CrewAI to develop AI agents and coordinate tasks among those agents. This structure enables systematic management of complex AI workflows while maintaining oversight of agent interactions and outcomes. The framework has the following components, which are shown in the following figure:

CrewAI Framework is built around the following components:

• Agents in CrewAI are autonomous components designed to perform specific tasks or roles within a multi-agent system. They have specific roles (such as researcher or writer) and make autonomous decisions with or without using external tools. LLMs are the core intelligence behind CrewAI agents. LLMs enable agents to understand context, make decisions, and generate human-like responses.
• Tasks are defined jobs assigned to agents with clear objectives, including execution details and required resources.
• Crews are coordinated teams of agents working together on a shared goal. Crews require defining agent roles, task assignments, and execution order.
• Tools refer to the skills or functions that agents can use to carry out various actions.
• Processes are responsible for orchestrating how tasks are executed by agents, similar to project management in human teams. These processes make sure that tasks are allocated and completed efficiently, in accordance with a predefined strategy.

Prerequisites

Before getting started with the solution, you need to get access to Amazon Bedrock models:

1. Sign in to the Amazon Bedrock console and in the navigation pane under Bedrock configurations, select Model access to request access to Amazon Bedrock models. This step is shown in the following screenshots.

In this example, we use Amazon Nova Pro through Amazon Bedrock as our LLM. CrewAI provides built-in integration with Amazon Bedrock.

2. Clone the GitHub repo into a local folder

git clone https://github.com/aws-samples/sample-compliance-assistant-with-agents.git

3. Use the following command to install the dependencies for running CrewAI in your Python environment:

pip install crewai uv

Your compliance agents

In this step, you will define your agents:

6. Define compliance agents in the agents.yaml file. Each agent has a specific role to play:

   compliance_analyst:
     role: {topic} Senior Compliance Analyst
     goal: Review and understand regulatory and compliance requirements around {topic}
     backstory: You're a seasoned Compliance analyst with deep expertise in areas such as PCI DSS, HIPAA, NIST, ISO and knack for uncovering the latest regulations and requirements in {topic}.
   compliance_specialist:
     role: {topic} Compliance Specialist
     goal: Create detailed reports based on {topic} compliance analysis and research findings
     backstory: You're a meticulous compliance specialist with deep understanding of compliance and regulatory landscape for Financial services and Technology Industry. You create standards and policies for the organization to meet regulations and compliance needs.

7. Define tasks for the agents:

   compliance_analysis_task:
     description: Conduct a thorough analysis about {topic}. Make sure you find relevant information given the current year is {current_year}.
     expected_output: A list with 10 bullet points of the most relevant information about {topic}
     agent: compliance_analyst
   compliance_reporting_task:
     description: Review the context you got and expand each topic into a full section for a report.
       Make sure the report is detailed and contains any and all relevant information for Financial Services Organization
     expected_output: A fully fledged report with the main topics, each with a full section of information.
     agent: compliance_specialist

8. The execution and process steps are defined in crew.py:

   def crew(self) -> Crew:
   """Creates the Compliance Automation crew"""
       return Crew(
          agents=self.agents, 
          tasks=self.tasks, 
          process=Process.sequential,
          verbose=True)

9. Define your LLM, topic, and runtime parameters in the .env file:

   MODEL=bedrock/us.amazon.nova-pro-v1:0
   AWS_REGION_NAME=us-west-2
   TOPIC='GDPR requirements for Data Privacy'

10. Run the crew as follows:

    crewai run

11. The following demo shows the output of the crew. You can see the agents collaborating to generate a detailed solution

In the output, notice that the compliance analyst and the compliance specialist are working together to solve multiple aspects of General Data Protection Regulation (GDPR) requirements for trading services. Note the synergistic collaboration between agents as they refine their approach and develop a comprehensive compliance management response through iterative problem-solving.

Addressing LLM challenges with domain-specific knowledge

LLMs, although impressive in their broad knowledge, face two key limitations when dealing with specialized domains or recent information. First, they struggle with specialized information specific to your organization. Second, because their knowledge is limited to their training data, they might not reflect the latest updates in rapidly changing fields. This limitation becomes particularly important when dealing with evolving compliance requirements, such as Payment Card Industry Data Security Standard (PCI DSS), GDPR, AML rules, and Know Your Customer (KYC) regulations. Additionally, organizations need solutions that are customized to their specific compliance requirements and internal standards, rather than generic responses from LLMs.

Retrieval Augmented Generation (RAG) is a technique that enables generative AI models to retrieve and incorporate current organizational and domain-specific information from external databases. Amazon Bedrock Knowledge Base is a managed capability that helps you implement the entire RAG technique without having to build custom integrations to data sources and manage data flows. By incorporating a knowledge base containing the latest publications, regulatory updates, and compliance guidelines from authoritative sources such as NIST, ISO, PCI, and regulatory bodies, Amazon Bedrock Knowledge Bases helps make sure that your AI system stays current with the latest compliance requirements. During prompt generation, RAG first retrieves relevant data from this continually updated knowledge base, then uses it to create informed responses. This helps provide more relevant, accurate, and customized responses aligned with current regulatory and organizational standards. For example, when querying about PCI DSS v4.0 requirements or recent GDPR amendments, the system can pull the most up-to-date information directly from authoritative sources rather than relying on potentially outdated training data.

Create an Amazon Bedrock knowledge base with contextual information from your data sources

1. From the Amazon Bedrock navigation pane, select Knowledge Bases under Builder tools and choose a Knowledge Base with vector store.
   
2. Provide the Knowledge Base name and Data source details. You’ll use the web crawler for ingesting data.
   
3. The web crawler provided by Amazon Bedrock connects to and crawls URLs you selected for use in your Amazon Bedrock knowledge base. Add the URLs as data sources under Source URLs.
   

4. Select the model for embeddings. We have selected Amazon Titan Text Embeddings v2, as shown in the following screenshot.
   

After a few minutes, the knowledge base will be ready. After it has synced, Amazon Bedrock Knowledge Bases handles generating, running, and formatting the result of the query, simplifying building natural language interfaces to structured data.

Amazon Bedrock Agents

Amazon Bedrock Agents is a comprehensive environment for building sophisticated AI agent systems. At its core, it enables seamless multi-agent collaboration and maintains conversation context through native memory retention across interactions. Amazon Bedrock Agents integrates naturally with knowledge bases and enforce security through built-in guardrails. For this solution, we focus on two key capabilities: the RAG feature, which allows agents to access and utilize information from knowledge bases, and the security features provided through Amazon Bedrock Guardrails. These guardrails serve as an essential safeguard for your generative AI applications, promoting responsible and secure AI interactions.

1. To create an agent, from the Amazon Bedrock navigation pane under Builder tools, select Agents and select Create Agent.
   

2. Under Agent details, choose the model. We use Amazon Nova Pro for our use case, as shown in the following screenshot.
   

3. Under Knowledge Bases, add knowledge bases to your agent.
   

4. Choose the knowledge base name from the dropdown list, as shown in the following screenshot.
   

Amazon Bedrock Guardrails

Amazon Bedrock Guardrails provides safety controls that help maintain responsible AI use by providing a layer of security. Guardrails provide content filtering to monitor and filter AI model outputs to help prevent harmful, inappropriate, or biased content. You can set up filters for things such as hate speech, explicit content, or personally identifiable information (PII). You can also apply customizable rules and input/output validation.

1. You can find Guardrails in the Amazon Bedrock navigation pane under Safeguards. Choose Create guardrail and provide a guardrail name
   

3. As shown in the following screenshot, select the content filters you want to implement for your Amazon Bedrock based application
   
4. Add denied topics with specific examples
   
5. After you’ve created your guardrail, attach guardrail to the agent.

Putting it all together: Integrating Amazon Bedrock Agents with CrewAI

CrewAI provides seamless integration with Amazon Bedrock features, including Amazon Bedrock Knowledge Bases and Amazon Bedrock Agents through CrewAI tools functionality. When these tools are triggered from CrewAI agents, they process your query, retrieve the relevant information from the Amazon Bedrock knowledge base, and return responses back to CrewAI agent.

1. Refer to the sample code demonstrating CrewAI tools for Amazon Bedrock Agent. You need to define your Amazon Bedrock AgentId and Alias as parameters in the .env file
2. Execute the crew again with Amazon Bedrock Agents:

   crewai run

3. You can find the generated output below
   

When you execute the crew, the compliance analyst agent initiates the process by invoking the CrewAI Bedrock tool to extract regulatory requirements from Amazon Bedrock Knowledge Bases, which is then seamlessly transformed into technical requirements by the compliance specialist agent. Through iterative collaboration, these specialized agents work together to fill information gaps, and the enterprise architect agent synthesizes the gathered insights to develop a robust implementation strategy and execution plan. This streamlined process demonstrates how multiple AI agents can effectively coordinate to transform compliance requirements into actionable technical solutions.

Clean up

To avoid ongoing charges, follow these steps to clean up resources:

1. Delete the Amazon Bedrock knowledge base that you created:

aws bedrock-agent delete-knowledge-base --knowledge-base-id <your-kb-id>

2. Delete the Amazon Bedrock agents that you created:

aws bedrock-agent delete-agent --agent-id <your-agent-id>

Conclusion

In this post, we demonstrated how to:

• Build a multi-agent AI system using CrewAI that mimics the structure of a comprehensive compliance department with specialized agents for different functions
• Enhance AI responses with domain-specific knowledge by implementing RAG using Amazon Bedrock Knowledge Bases
• Safeguard your generative AI applications with Amazon Bedrock Guardrails to help prevent harmful, inappropriate, or biased content
• Create custom tools in CrewAI to integrate with Amazon Bedrock Agents for more powerful and context-aware compliance solutions
• Automate the entire compliance lifecycle from monitoring regulatory changes to implementing technical controls without extensive manual effort
• Deploy a production-ready solution that continually adapts to evolving regulatory requirements in financial services and other highly regulated industries

This solution combines Amazon Bedrock Knowledge Bases and CrewAI to create smart, multi-agent AI systems that help streamline regulatory compliance tasks. With simplified RAG implementation, sophisticated workflows that mirror human teams, and faster adaptation to new regulations, this approach shows how AI can assist organizations with specific aspects of complex regulatory requirements.

This solution serves as a practical starting point for organizations looking to enhance their compliance processes with AI capabilities, demonstrating how intelligent systems could complement and streamline existing compliance workflows. The complete source code for this project is available on the GitHub repository. Feel free to explore, fork, or contribute!

About the Authors

Balu Mathew is a Senior Solutions Architect at AWS, based in Raleigh, NC. He collaborates with Global Financial Services customers to design and implement secure, scalable and resilient solutions on AWS. With deep expertise in security, machine learning, and the financial services industry, he helps organizations build, protect, and scale large-scale distributed systems efficiently. Outside of work, he enjoys spending time with his kids and exploring the mountains and the outdoors.

Read More