Month: March 2025

Researchers and enthusiasts have been fascinated by the challenge of reverse-engineering complex behaviors that emerge from simple rules in cellular automata for decades. Traditionally, this field takes a bottom-up approach—defining local regulations and observing the patterns arising from them. But what if we could flip this process? Instead of manually designing rules, we could develop a fully differentiable system that learns the local rules necessary to generate a given complex pattern while maintaining the discrete nature of cellular automata. This approach opens new possibilities for automating rule discovery in a structured and scalable way.

Previous work has investigated learning transition rules using non-differentiable methods, proving that this method can evolve local regulations for specific computational tasks. Additionally, research has explored ways to make one-dimensional cellular automata differentiable, enabling gradient-based optimization techniques for rule learning. Building on these foundations allows us to develop systems that automatically discover rules that generate desired patterns, bridging the gap between handcrafted cellular automata and learned computational models.

Google researchers introduced Differentiable Logic Cellular Automata (DiffLogic CA), which applies differentiable logic gates to cellular automata. This method successfully replicates the rules of Conway’s Game of Life and generates patterns through learned discrete dynamics. The approach merges Neural Cellular Automata (NCA), which can learn arbitrary behaviors but lack discrete state constraints, with Differentiable Logic Gate Networks, which enable combinatorial logic discovery but have not been tested in recurrent settings. This integration paves the way for learnable, local, and discrete computing, potentially advancing programmable matter. The study explores whether Differentiable Logic CA can learn and generate complex patterns akin to traditional NCAs.

NCA integrates classical cellular automata with deep learning, enabling self-organization through learnable update rules. Unlike traditional methods, NCA uses gradient descent to discover dynamic interactions while preserving locality and parallelism. A 2D grid of cells evolves via perception (using Sobel filters) and update stages (through neural networks). Differentiable Logic Gate Networks (DLGNs) extend this by replacing neurons with logic gates, allowing discrete operations to be learned via continuous relaxations. DiffLogic CA further integrates these concepts, employing binary-state cells with logic gate-based perception and update mechanisms, forming an adaptable computational system akin to programmable matter architectures like CAM-8.

Conway’s Game of Life, a cellular automaton introduced by John Conway in 1970, follows simple rules governing cell interactions to produce complex behaviors. A model was trained using DiffLogic CA to replicate these rules, employing a network with 16 perception circuit-kernels and 23 update layers. The loss function minimized squared differences between predicted and actual states. Training on all 512 possible 3×3 grids enabled accurate rule learning, which scaled effectively to larger grids. The learned circuit replicated classic Game of Life patterns, demonstrating its ability to generalize, exhibit fault tolerance, and self-heal without explicitly designed robustness mechanisms.

In conclusion, the study introduces DiffLogic CA, a NCA architecture that employs discrete cell states and recurrent binary circuits. Integrating Deep Differentiable Logic Networks enables the differentiable training of logic gates. The model replicates Conway’s Game of Life and generates patterns using learned discrete dynamics. Unlike traditional NCAs, which rely on costly matrix operations, this approach enhances interpretability and efficiency. Future improvements may involve hierarchical architectures and LSTM-like gating mechanisms. This research suggests that integrating differentiable logic gates with NCAs could advance programmable matter, making computation more efficient and adaptable to complex pattern generation.

Check out the Technical details. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Meet Parlant: An LLM-first conversational AI framework designed to provide developers with the control and precision they need over their AI customer service agents, utilizing behavioral guidelines and runtime supervision. It’s operated using an easy-to-use CLI and native client SDKs in Python and TypeScript .
The post Google AI Introduces Differentiable Logic Cellular Automata (DiffLogic CA): A Differentiable Logic Approach to Neural Cellular Automata appeared first on MarkTechPost.

Kaggle Kernels (also called Notebooks) represent a revolutionary cloud-based platform for data science and machine learning work. They provide a complete computational environment where you can write, run, and visualize code directly in your browser without any local setup or installation.

What makes Kaggle Kernels particularly valuable:

Zero configuration required: Everything is pre-installed and ready to use immediately

Free access to powerful computing resources: CPUs, GPUs, and TPUs available at no cost

Browser-based accessibility: Work from any device with an internet connection

Integrated ecosystem: Seamless access to datasets, competitions, and community resources

Reproducible research: Complete environment captured in shareable documents

Collaborative features: Learn from others and share your own work

This tutorial will guide you through everything you need to know about Kaggle Kernels, from account setup to developing sophisticated machine learning models.

Prerequisites

A web browser (Chrome, Firefox, Safari, or Edge)

Basic understanding of Python or R (though beginners can still follow along)

Interest in data science and machine learning

1. Creating and Setting Up Your Kaggle Account

Sign-Up Process

Navigate to www.kaggle.com

Click the “Register” button in the top-right corner

Choose to sign up with Google, Facebook, or email credentials

Complete your profile with a username, profile picture, and bio

Verify your email address through the confirmation link

2. Navigating the Kaggle Platform

Understanding the Interface

The Kaggle platform has several key sections accessed through the top navigation bar:

Home: Personalized feed of activity and recommendations

Competitions: Active and past machine learning competitions

Datasets: Repository of public datasets to explore and use

Models: Space to explore and use different models

Code: Where you access Notebooks (formerly Kernels)

Discussion: Community forums and conversations

Learn: Educational courses on data science and ML

Accessing Notebooks/Kernels

Click on “Code” in the top navigation bar

You’ll see a page with featured notebooks and your own work

Click on “New Notebook” button to create a new notebook

3. Creating Your First Kernel

Click the “New Notebook” button, this will open up a fresh notebook 

The Kaggle Kernel environment has several key components:

Code Editor: Where you write your Python/R code

Output Area: Displays results, plots, and print statements

File Browser: Access datasets and output files

Settings Panel: Configure hardware accelerators and other options

5. Adding Data to Your Kernel

There are three ways to add data:

From Kaggle Datasets:

Click “Add Input” in the top-right corner

Search for and select a dataset

Click “Add” to include it in your project

From a Competition:

If you created a kernel from a competition, the data is already available

Access it in the /kaggle/input/ directory

Upload Your Own Data:

Click “Add data” > “Upload”

Select files from your computer (max 20GB)

6. Writing and Running Code

Type your code in a code cell

Press “Shift+Enter” or click the “Run” button to execute

Add a new cell by clicking “+” or pressing “Esc+B”

Change cell type (code/markdown) using the dropdown in the toolbar

Example: Loading Data and Creating a Simple Model

7. Using GPU/TPU Accelerators

For deep learning and resource-intensive tasks:

Click on the “Settings” tab

Under “Accelerator”, select:

None (default CPU)

GPU (T4 x2)

GPU P100

TPU VM (v3-8)

Save your settings

8. Installing Additional Packages

You can install additional packages using pip:

Or add them to the settings:

Go to “Add-ons” > “Install Dependencies”

It shall open a side window

Enter the package name and version (optional)

9. Saving and Sharing Your Work

Save Version:

Click “Save Version” to create a snapshot

Add a version name and description

Choose visibility (Public/Private)

Share Your Kernel:

Click “Share” button in the top-right

Get a shareable link or publish to the Kaggle community

10. Forking and Collaborating

To build upon someone else’s work:

Find a public notebook you like

Click “Copy & Edit” to create your own version

Make changes and save your version

11. Common Keyboard Shortcuts

For faster workflow:

Shift+Enter: Run current cell

Ctrl+Enter: Run current cell without moving to next cell

Alt+Enter: Run current cell and insert new cell below

Esc+A: Insert cell above

Esc+B: Insert cell below

Esc+D,D: Delete current cell

Esc+M: Change to Markdown cell

Esc+Y: Change to Code cell

12. Troubleshooting

Common issues and solutions:

Kernel Timeouts:

Sessions automatically terminate after 9 hours of inactivity

Save your work frequently

Memory Errors:

Reduce data size or batch processing

Use more efficient algorithms/data structures

Package Installation Errors:

Check for compatibility issues

Try different versions of packages

Conclusion

Kaggle Kernels provide an excellent environment for learning and experimenting with machine learning. You can access powerful computational resources for free, collaborate with others, and participate in competitions to sharpen your skills.

Next Steps

Explore the Kaggle Learn platform for tutorials

Join a competition to apply your skills

Study public notebooks to learn from the community

Share your own work to get feedback

Happy coding and machine learning!
The post Getting Started with Kaggle Kernels for Machine Learning appeared first on MarkTechPost.

The landscape of generative AI and LLMs has experienced a remarkable leap forward with the launch of Mercury by the cutting-edge startup Inception Labs. Introducing the first-ever commercial-scale diffusion large language models (dLLMs), Inception labs promises a paradigm shift in speed, cost-efficiency, and intelligence for text and code generation tasks.

Mercury: Setting New Benchmarks in AI Speed and Efficiency

Inception’s Mercury series of diffusion large language models introduces unprecedented performance, operating at speeds previously unachievable with traditional LLM architectures. Mercury achieves remarkable throughput—over 1000 tokens per second on commodity NVIDIA H100 GPUs—a performance that was formerly exclusive to custom-designed hardware like Groq, Cerebras, and SambaNova. This translates to an astonishing 5-10x speed increase compared to current leading autoregressive models.

Diffusion Models: The Future of Text Generation

Traditional autoregressive LLMs generate text sequentially, token-by-token, causing significant latency and computational costs, especially in extensive reasoning and error-correction tasks. Diffusion models, however, leverage a unique “coarse-to-fine” generation process. Unlike autoregressive models restricted by sequential generation, diffusion models iteratively refine outputs from noisy approximations, enabling parallel token updates. This method significantly enhances reasoning, error correction, and overall coherence of the generated content.

While diffusion approaches have proven revolutionary in image, audio, and video generation—powering applications like Midjourney and Sora—their application in discrete data domains such as text and code was largely unexplored until Inception’s breakthrough.

Mercury Coder: High-Speed, High-Quality Code Generation

Inception’s flagship product, Mercury Coder, is optimized specifically for coding applications. Developers now have access to a high-quality, rapid-response model capable of generating code at more than 1000 tokens per second, a dramatic improvement over existing speed-focused models.

On standard coding benchmarks, Mercury Coder doesn’t just match but often surpasses the performance of other high-performing models such as GPT-4o Mini and Claude 3.5 Haiku. Moreover, Mercury Coder Mini secured a top-ranking position on Copilot Arena, tying for second place and outperforming established models like GPT-4o Mini and Gemini-1.5-Flash. Even more impressively, Mercury accomplishes this while maintaining approximately 4x faster speeds than GPT-4o Mini.

Versatility and Integration

Mercury dLLMs function seamlessly as drop-in replacements for traditional autoregressive LLMs. They effortlessly support use-cases including Retrieval-Augmented Generation (RAG), tool integration, and agent-based workflows. The diffusion model’s parallel refinement allows multiple tokens to be updated simultaneously, ensuring swift and accurate generation suitable for enterprise environments, API integration, and on-premise deployments.

Built by AI Innovators

Inception’s technology is underpinned by foundational research at Stanford, UCLA and Cornell from its pioneering founders, recognized for their crucial contributions to the evolution of generative AI. Their combined expertise includes the original development of image-based diffusion models and innovations such as Direct Preference Optimization, Flash Attention, and Decision Transformers—techniques widely acknowledged for their transformative impact on modern AI.

Inception’s introduction of Mercury marks a pivotal moment for enterprise AI, unlocking previously impossible performance levels, accuracy, and cost-efficiency.

Check out the Playground and Technical details. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Inception Unveils Mercury: The First Commercial-Scale Diffusion Large Language Model appeared first on MarkTechPost.

Researchers at The Ohio State University have introduced Finer-CAM, an innovative method that significantly improves the precision and interpretability of image explanations in fine-grained classification tasks. This advanced technique addresses key limitations of existing Class Activation Map (CAM) methods by explicitly highlighting subtle yet critical differences between visually similar categories.

Current Challenge with Traditional CAM

Conventional CAM methods typically illustrate general regions influencing a neural network’s predictions but frequently fail to distinguish fine details necessary for differentiating closely related classes. This limitation poses significant challenges in fields requiring precise differentiation, such as species identification, automotive model recognition, and aircraft type differentiation.

Finer-CAM: Methodological Breakthrough

The central innovation of Finer-CAM lies in its comparative explanation strategy. Unlike traditional CAM methods that focus solely on features predictive of a single class, Finer-CAM explicitly contrasts the target class with visually similar classes. By calculating gradients based on the difference in prediction logits between the target class and its similar counterparts, it reveals unique image features, enhancing the clarity and accuracy of visual explanations.

Finer-CAM Pipeline

The methodological pipeline of Finer-CAM involves three main stages:

Feature Extraction:

An input image first passes through neural network encoder blocks, generating intermediate feature maps.

A subsequent linear classifier uses these feature maps to produce prediction logits, which quantify the confidence of predictions for various classes.

Gradient Calculation (Logit Difference):

Standard CAM methods calculate gradients for a single class.

Finer-CAM computes gradients based on the difference between the prediction logits of the target class and a visually similar class.

This comparison identifies the subtle visual features specifically discriminative to the target class by suppressing commonly shared features.

Activation Highlighting:

The gradients calculated from the logit difference are used to produce enhanced class activation maps that emphasize discriminative visual details crucial for distinguishing between similar categories.

Experimental Validation

B.1. Model Accuracy

Researchers evaluated Finer-CAM across two popular neural network backbones, CLIP and DINOv2. Experiments demonstrated that DINOv2 generally produces higher-quality visual embeddings, achieving superior classification accuracy compared to CLIP across all tested datasets.

B.2. Results on FishVista and Aircraft

Quantitative evaluations on the FishVista and Aircraft datasets further demonstrate Finer-CAM’s effectiveness. Compared to baseline CAM methods (Grad-CAM, Layer-CAM, Score-CAM), Finer-CAM consistently delivered improved performance metrics, notably in relative confidence drop and localization accuracy, underscoring its ability to highlight discriminative details crucial for fine-grained classification.

B.3. Results on DINOv2

Additional evaluations using DINOv2 as the backbone showed that Finer-CAM consistently outperformed baseline methods. These results indicate that Finer-CAM’s comparative method effectively enhances localization performance and interpretability. Due to DINOv2’s high accuracy, more pixels need to be masked to significantly impact predictions, resulting in larger deletion AUC values and occasionally smaller relative confidence drops compared to CLIP.

Visual and Quantitative Advantages

Highly Precise Localization: Clearly pinpoints discriminative visual features, such as specific coloration patterns in birds, detailed structural elements in cars, and subtle design variations in aircraft.

Reduction of Background Noise: Significantly reduces irrelevant background activations, increasing the relevance of explanations.

Quantitative Excellence: Outperforms traditional CAM approaches (Grad-CAM, Layer-CAM, Score-CAM) in metrics including relative confidence drop and localization accuracy.

Extendable to multi-modal zero-shot learning scenarios

Finer-CAM is extendable to multi-modal zero-shot learning scenarios. By intelligently comparing textual and visual features, it accurately localizes visual concepts within images, significantly expanding its applicability and interpretability.

Researchers have made Finer-CAM’s source code and colab demo available.

Check out the Paper, Github and Colab demo. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Finer-CAM Revolutionizes AI Visual Explainability: Unlocking Precision in Fine-Grained Image Classification appeared first on MarkTechPost.

Large Language Models (LLMs) benefit significantly from reinforcement learning techniques, which enable iterative improvements by learning from rewards. However, training these models efficiently remains challenging, as they often require extensive datasets and human supervision to enhance their capabilities. Developing methods that allow LLMs to self-improve autonomously without additional human input or large-scale architectural modifications has become a major focus in AI research.

The key challenge in training LLMs is ensuring the learning process is efficient and structured. The training process can stall when models encounter problems beyond their capabilities, leading to poor performance. Traditional reinforcement learning techniques rely on well-curated datasets or human feedback to create effective learning pathways, but this approach is resource-intensive. Also, LLMs struggle to improve systematically without a structured difficulty gradient, making it difficult to bridge the gap between basic reasoning tasks and more complex problem-solving.

Existing approaches to training LLMs primarily involve supervised fine-tuning, reinforcement learning from human feedback (RLHF), and curriculum learning. Supervised fine-tuning requires manually labeled datasets, which can lead to overfitting and limited generalization. RLHF introduces a layer of human oversight, where models are refined based on human evaluations, but this method is costly and does not scale efficiently. Curriculum learning, which gradually increases task difficulty, has shown promise, but current implementations still rely on pre-defined datasets rather than allowing models to generate their learning trajectories. These limitations highlight the need for an autonomous learning framework that enables LLMs to improve their problem-solving abilities independently.

Researchers from Tufa Labs introduced LADDER (Learning through Autonomous Difficulty-Driven Example Recursion) to overcome these limitations. This framework enables LLMs to self-improve by recursively generating and solving progressively simpler variants of complex problems. Unlike prior methods that depend on human intervention or curated datasets, LADDER leverages the model’s capabilities to create a natural difficulty gradient, allowing for structured self-learning. The research team developed and tested LADDER on mathematical integration tasks, demonstrating its effectiveness in enhancing model performance. By applying LADDER, the researchers enabled a 3-billion-parameter Llama 3.2 model to improve its accuracy on undergraduate integration problems from 1% to 82%, an unprecedented leap in mathematical reasoning capabilities. Also, the approach was extended to larger models, such as Qwen2.5 7B Deepseek-R1 Distilled, achieving 73% accuracy on the MIT Integration Bee qualifying examination, far surpassing models like GPT-4o, which gained only 42%, and typical human performance in the 15-30% range.

LADDER follows a structured methodology that allows LLMs to bootstrap their learning by systematically breaking down complex problems. The process involves three primary components: variant generation, solution verification, and reinforcement learning. The variant generation step ensures the model produces progressively easier versions of a given problem, forming a structured difficulty gradient. The solution verification step employs numerical integration methods to assess the correctness of generated solutions, providing immediate feedback without human intervention. Finally, the reinforcement learning component uses Group Relative Policy Optimization (GRPO) to train the model efficiently. This protocol enables the model to learn incrementally by leveraging verified solutions, allowing it to refine its problem-solving strategies systematically. The researchers extended this approach with Test-Time Reinforcement Learning (TTRL), which dynamically generates problem variants during inference and applies reinforcement learning to refine solutions in real time. When applied to the MIT Integration Bee qualifying examination, TTRL boosted model accuracy from 73% to 90%, surpassing OpenAI’s o1 model.

When tested on a dataset of 110 undergraduate-level integration problems, a Llama 3.2 3B model trained with LADDER achieved 82% accuracy, compared to 2% accuracy when using pass@10 sampling. The approach also demonstrated scalability, as increasing the number of generated variants led to continued performance improvements. In contrast, reinforcement learning without variants failed to achieve meaningful gains, reinforcing the importance of structured problem decomposition. The researchers observed that LADDER-trained models could solve integrals requiring advanced techniques that were previously out of reach. Applying the methodology to the MIT Integration Bee qualifying examination, a Deepseek-R1 Qwen2.5 7B model trained with LADDER outperformed larger models that did not undergo recursive training, showcasing the effectiveness of structured self-improvement in mathematical reasoning.

Key Takeaways from the Research on LADDER include:

Enables LLMs to self-improve by recursively generating and solving simpler variants of complex problems.

Llama 3.2 3B model improved from 1% to 82% on undergraduate integration tasks, demonstrating the effectiveness of structured self-learning.

Qwen2.5 7B Deepseek-R1 Distilled achieved 73% accuracy, outperforming GPT-4o (42%) and exceeding human performance (15-30%).

Further boosted accuracy from 73% to 90%, surpassing OpenAI’s o1 model.

LADDER does not require external datasets or human intervention, making it a cost-effective and scalable solution for LLM training.

Models trained with LADDER demonstrated superior problem-solving capabilities compared to reinforcement learning without structured difficulty gradients.

The framework provides a structured way for AI models to refine their reasoning skills without external supervision.

The methodology can be extended to competitive programming, theorem proving, and agent-based problem-solving.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Tufa Labs Introduced LADDER: A Recursive Learning Framework Enabling Large Language Models to Self-Improve without Human Intervention appeared first on MarkTechPost.

Deep learning faces difficulties when applied to large physical systems on irregular grids, especially when interactions occur over long distances or at multiple scales. Handling these complexities becomes harder as the number of nodes increases. Several techniques have difficulty tackling these big problems, resulting in high computational costs and inefficiency. Some major issues are capturing long-range effects, handling multi-scale dependencies, and efficient computation with minimal resource usage. These issues make it difficult to apply deep learning models effectively to fields like molecular simulations, weather prediction, and particle mechanics, where large datasets and complex interactions are common.

Currently, Deep learning methods struggle with scaling attention mechanisms for large physical systems. Traditional self-attention computes interactions between all points, leading to extremely high computational costs. Some methods apply attention to small patches, like SwinTransformer for images, but irregular data needs extra steps to structure it. Techniques like PointTransformer use space-filling curves, but this can break spatial relationships. Hierarchical methods, such as H-transformer and OctFormer, group data at different levels but rely on costly operations. Cluster attention methods reduce complexity by aggregating points, but this process loses fine details and struggles with multi-scale interactions.

To address these problems, researchers from AMLab, University of Amsterdam and CuspAI introduced Erwin, a hierarchical transformer that enhances data processing efficiency through ball tree partitioning. The attention mechanism enables parallel computation across clusters through ball tree partitions that partition data hierarchically to structure its computations. This approach minimizes computational complexity without sacrificing accuracy, bridging the gap between the efficiency of tree-based methods and the generality of attention mechanisms. Erwin uses self-attention in localized regions with positional encoding and distance-based attention bias to capture geometric structures. Cross-ball connections facilitate communication among various sections, with tree coarsening and refinement mechanisms balancing global and local interactions. Scalability and expressivity with minimal computational expense are guaranteed through this organized process.

Researchers conducted experiments to evaluate Erwin. It outperformed equivariant and non-equivariant baselines in cosmological simulations, capturing long-range interactions and improving with larger training datasets. For molecular dynamics, it accelerated simulations by 1.7–2.5 times without compromising accuracy, surpassing MPNN and PointNet++ in runtime while maintaining competitive test loss. Erwin outperformed MeshGraphNet, GAT, DilResNet, and EAGLE in turbulent fluid dynamics, excelling in pressure prediction while being three times faster and using eight times less memory than EAGLE. Larger ball sizes in cosmology enhanced performance by retaining long-range dependencies but increased the computational runtime, and applying MPNN at the embedding step improved the local interactions in molecular dynamics.

The hierarchical transformer design proposed here effectively handles large-scale physical systems with ball tree partitioning and obtains state-of-the-art cosmology and molecular dynamics results. Although its optimized structure compromises between expressivity and runtime, it has computational overhead from padding and high memory requirements. Future work can investigate learnable pooling and other geometric encoding strategies to enhance efficiency. Erwin’s performance and scalability in all domains make it a reference point for developments in modeling large particle systems, computational chemistry, and molecular dynamics.

Check out the Paper and GitHub Page. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Researchers from AMLab and CuspAI Introduced Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems appeared first on MarkTechPost.

Transformer models have transformed language modeling by enabling large-scale text generation with emergent properties. However, they struggle with tasks that require extensive planning. Researchers have explored modifications in architecture, objectives, and algorithms to improve their ability to achieve goals. Some approaches move beyond traditional left-to-right sequence modeling by incorporating bidirectional context, as seen in models trained on past and future information. Others attempt to optimize the generation order, such as latent-variable modeling or binary tree-based decoding, though left-to-right autoregressive methods often remain superior. A more recent approach involves jointly training a transformer for forward and backward decoding, enhancing the model’s ability to maintain compact belief states.

Further research has explored predicting multiple tokens simultaneously to improve efficiency. Some models have been designed to generate more than one token at a time, leading to faster and more robust text generation. Pretraining on multi-token prediction has been shown to enhance large-scale performance. Another key insight is that transformers encode belief states non-compactly within their residual stream. In contrast, state-space models offer more compact representations but come with trade-offs. For instance, certain training frameworks struggle with specific graph structures, revealing limitations in existing methodologies. These findings highlight ongoing efforts to refine transformer architectures for better structured and efficient sequence modeling.

Researchers from Microsoft Research, the University of Pennsylvania, UT Austin, and the University of Alberta introduced the Belief State Transformer (BST). This model enhances next-token prediction by considering both prefix and suffix contexts. Unlike standard transformers, BST encodes information bidirectionally, predicting the next token after the prefix and the previous token before the suffix. This approach improves performance on challenging tasks, such as goal-conditioned text generation and structured prediction problems like star graphs. By learning a compact belief state, BST outperforms conventional methods in sequence modeling, offering more efficient inference and stronger text representations, with promising implications for large-scale applications.

Unlike traditional next-token prediction models, the BST is designed to enhance sequence modeling by integrating both forward and backward encoders. It utilizes a forward encoder for prefixes and a backward encoder for suffixes, predicting the next and previous tokens. This approach prevents models from adopting shortcut strategies and improves long-term dependency learning. BST outperforms baselines in star graph navigation, where forward-only Transformers struggle. Ablations confirm that the belief state objective and backward encoder are essential for performance. During inference, BST omits the backward encoder, maintaining efficiency while ensuring goal-conditioned behavior.

Unlike forward-only and multi-token models, the BST effectively constructs a compact belief state. A belief state encodes all necessary information for future predictions. The BST learns such representations by jointly modeling prefixes and suffixes, enabling goal-conditioned text generation. Experiments using TinyStories show BST outperforms the Fill-in-the-Middle (FIM) model, producing more coherent and structured narratives. Evaluation with GPT-4 reveals BST’s superior storytelling ability, with clearer connections between prefix, generated text, and suffix. Additionally, BST excels in unconditional text generation by selecting sequences with high-likelihood endings, demonstrating its advantages over traditional next-token predictors.

In conclusion, the BST improves goal-conditioned next-token prediction by addressing the limitations of traditional forward-only models. It constructs a compact belief state, encoding all necessary information for future predictions. Unlike conventional transformers, BST predicts the next token for a prefix and the previous token for a suffix, making it more effective in complex tasks. Empirical results demonstrate its advantages in story writing, outperforming the Fill-in-the-Middle approach. While our experiments validate its performance on small-scale tasks, further research is needed to explore its scalability and applicability to broader goal-conditioned problems, enhancing efficiency and inference quality.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Microsoft AI Introduces Belief State Transformer (BST): Enhancing Goal-Conditioned Sequence Modeling with Bidirectional Context appeared first on MarkTechPost.

Large language models have made significant strides in understanding and generating human-like text. Yet, when it comes to complex reasoning tasks—especially those that require multi-step calculations or logical analysis—they often struggle. Traditional chain-of-thought (CoT) approaches help by breaking down problems into intermediate steps, but they rely heavily on the model’s internal reasoning. This internal dependency can sometimes lead to mistakes, particularly with intricate computations or when multiple reasoning steps are needed. In such cases, minor errors may accumulate, resulting in outcomes that are not as precise as expected. The need for a method that can verify and adjust its own reasoning is clear, especially in tasks like scientific analysis or competition-level mathematics.

Researchers at Alibaba have proposed a new AI tool called START, which stands for Self-Taught Reasoner with Tools. Rather than relying solely on internal logic, START integrates an external Python interpreter to assist with reasoning tasks. The model is built on a fine-tuned version of the QwQ-32B model and employs a two-fold strategy to improve its problem-solving skills. First, it uses a method called Hint-infer. Here, the model is encouraged to include prompts like “Wait, maybe using Python here is a good idea,” which signal that it should perform computations or self-check its work using external tools. Second, the model undergoes a fine-tuning process known as Hint Rejection Sampling Fine-Tuning (Hint-RFT). This process refines the model’s reasoning by filtering and modifying its output based on how effectively it can invoke external tools. The result is a model that is not only capable of generating a logical chain of thought but also of verifying its steps through external computation.

Technical Insights and Benefits

At its core, START is an evolution of the chain-of-thought approach. Its two-stage training process is designed to help the model use external tools as a natural extension of its reasoning process. In the first stage, Hint-infer allows the model to integrate cues that prompt tool usage. These hints are strategically inserted at points where the model might be reconsidering its approach, often after transitional words like “Alternatively” or “Wait.” This encourages the model to verify its reasoning with Python code, leading to self-correction when necessary.

In the second stage, Hint-RFT takes the output generated with these hints and refines it. By scoring and filtering the reasoning steps, the model learns to better decide when and how to invoke external tools. The refined dataset from this process is then used to fine-tune the model further, resulting in a version of QwQ-32B that we now call START. The integration of external computation is a thoughtful addition that helps minimize errors, ensuring that the model’s reasoning is both coherent and more reliable.

Empirical Findings and Insights

The researchers evaluated START on a range of tasks, including graduate-level science questions, challenging math problems, and programming tasks. Across these domains, START showed notable improvements over its base model. For example, on a set of PhD-level science questions, the model achieved an accuracy of 63.6%, which is a modest yet meaningful improvement over the original model’s performance. On math benchmarks—ranging from high school level to competition problems—the accuracy improvements were similarly encouraging. These results suggest that the ability to incorporate external verification can lead to better problem-solving, especially in tasks where precision is crucial.

In programming challenges, START’s approach allowed it to generate and test code snippets, leading to a higher rate of correct solutions compared to models that rely solely on internal reasoning. Overall, the study indicates that the integration of tool usage within the reasoning process can help models produce more accurate and verifiable results.

Concluding Thoughts

The development of START offers a thoughtful step forward in addressing the inherent challenges of complex reasoning in large language models. By combining internal chain-of-thought reasoning with external tool integration, the model provides a practical solution to some of the persistent issues in computational and logical tasks. The approach is both simple and elegant: encouraging the model to self-check its work using an external Python interpreter and then fine-tuning it based on this ability leads to improved performance across diverse benchmarks.

This work is a promising example of how incremental refinements—in this case, the use of strategic hints and external computation—can significantly enhance the reliability of reasoning in language models. It demonstrates that by thoughtfully integrating external tools, we can guide models toward more accurate and reliable outcomes, especially in areas where precise computation and logical rigor are essential. The work behind START is an encouraging move toward models that are not only more capable but also more reflective and self-correcting in their approach to problem-solving.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, feel free to follow us on Twitter and don’t forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Alibaba Researchers Propose START: A Novel Tool-Integrated Long CoT Reasoning LLM that Significantly Enhances Reasoning Capabilities by Leveraging External Tools appeared first on MarkTechPost.

This post is co-authored with Sundeep Sardana, Malolan Raman, Joseph Lam, Maitri Shah and Vaibhav Singh from Verisk.
Verisk (Nasdaq: VRSK) is a leading strategic data analytics and technology partner to the global insurance industry, empowering clients to strengthen operating efficiency, improve underwriting and claims outcomes, combat fraud, and make informed decisions about global risks. Through advanced data analytics, software, scientific research, and deep industry knowledge, Verisk helps build global resilience across individuals, communities, and businesses. At the forefront of using generative AI in the insurance industry, Verisk’s generative AI-powered solutions, like Mozart, remain rooted in ethical and responsible AI use. Mozart, the leading platform for creating and updating insurance forms, enables customers to organize, author, and file forms seamlessly, while its companion uses generative AI to compare policy documents and provide summaries of changes in minutes, cutting the change adoption time from days or weeks to minutes.
The generative AI-powered Mozart companion uses sophisticated AI to compare legal policy documents and provides essential distinctions between them in a digestible and structured format. The new Mozart companion is built using Amazon Bedrock. Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models (FMs) from leading AI companies like AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon through a single API, along with a broad set of capabilities to build generative AI applications with security, privacy, and responsible AI. The Mozart application rapidly compares policy documents and presents comprehensive change details, such as descriptions, locations, excerpts, in a tracked change format.
The following screenshot shows an example of the output of the Mozart companion displaying the summary of changes between two legal documents, the excerpt from the original document version, the updated excerpt in the new document version, and the tracked changes represented with redlines.

In this post, we describe the development journey of the generative AI companion for Mozart, the data, the architecture, and the evaluation of the pipeline.
Data: Policy forms
Mozart is designed to author policy forms like coverage and endorsements. These documents provide information about policy coverage and exclusions (as shown in the following screenshot) and help in determining the risk and premium associated with an insurance policy.

Solution overview
The policy documents reside in Amazon Simple Storage Service (Amazon S3) storage. An AWS Batch job reads these documents, chunks them into smaller slices, then creates embeddings of the text chunks using the Amazon Titan Text Embeddings model through Amazon Bedrock and stores them in an Amazon OpenSearch Service vector database. Along with each document slice, we store the metadata associated with it using an internal Metadata API, which provides document characteristics like document type, jurisdiction, version number, and effective dates. This process has been implemented as a periodic job to keep the vector database updated with new documents. During the solution design process, Verisk also considered using Amazon Bedrock Knowledge Bases because it’s purpose built for creating and storing embeddings within Amazon OpenSearch Serverless. In the future, Verisk intends to use the Amazon Titan Embeddings V2 model.
The user can pick the two documents that they want to compare. This action invokes an AWS Lambda function to retrieve the document embeddings from the OpenSearch Service database and present them to Anthropic’s Claude 3 Sonnet FM, which is accessed through Amazon Bedrock. The results are stored in a JSON structure and provided using the API service to the UI for consumption by the end-user.
The following diagram illustrates the solution architecture.

Security and governance
Generative AI is very new technology and brings with it new challenges related to security and compliance. Verisk has a governance council that reviews generative AI solutions to make sure that they meet Verisk’s standards of security, compliance, and data use. Verisk also has a legal review for IP protection and compliance within their contracts. It’s important that Verisk makes sure the data that is shared by the FM is transmitted securely and the FM doesn’t retain any of their data or use it for its own training. The quality of the solution, speed, cost, and ease of use were the key factors that led Verisk to pick Amazon Bedrock and Anthropic’s Claude Sonnet within their generative AI solution.
Evaluation criteria
To assess the quality of the results produced by generative AI, Verisk evaluated based on the following criteria:

Accuracy
Consistency
Adherence to context
Speed and cost

To assess the generative AI results’ accuracy and consistency, Verisk designed human evaluation metrics with the help of in-house insurance domain experts. Verisk conducted multiple rounds of human evaluation of the generated results. During these tests, in-house domain experts would grade accuracy, consistency, and adherence to context on a manual grading scale of 1–10. The Verisk team measured how long it took to generate the results by tracking latency. Feedback from each round of tests was incorporated in subsequent tests.
The initial results that Verisk got from the model were good but not close to the desired level of accuracy and consistency. The development process underwent iterative improvements that included redesign, making multiple calls to the FM, and testing various FMs. The primary metric used to evaluate the success of FM and non-FM solutions was a manual grading system where business experts would grade results and compare them. FM solutions are improving rapidly, but to achieve the desired level of accuracy, Verisk’s generative AI software solution needed to contain more components than just FMs. To achieve the desired accuracy, consistency, and efficiency, Verisk employed various techniques beyond just using FMs, including prompt engineering, retrieval augmented generation, and system design optimizations.
Prompt optimization
The change summary is different than showing differences in text between the two documents. The Mozart application needs to be able to describe the material changes and ignore the noise from non-meaningful changes. Verisk created prompts using the knowledge of their in-house domain experts to achieve these objectives. With each round of testing, Verisk added detailed instructions to the prompts to capture the pertinent information and reduce possible noise and hallucinations. The added instructions would be focused on reducing any issues identified by the business experts reviewing the end results. To get the best results, Verisk needed to adjust the prompts based on the FM used—there are differences in how each FM responds to prompts, and using the prompts specific to the given FM provides better results. Through this process, Verisk instructed the model on the role it is playing along with the definition of common terms and exclusions. In addition to optimizing prompts for the FMs, Verisk also explored techniques for effectively splitting and processing the document text itself.
Splitting document pages
Verisk tested multiple strategies for document splitting. For this use case, a recursive character text splitter with a chunk size of 500 characters with 15% overlap provided the best results. This splitter is part of the LangChain framework; it’s a semantic splitter that considers semantic similarities in the text. Verisk also considered the NLTK splitter. With an effective approach for splitting the document text into processable chunks, Verisk then focused on enhancing the quality and relevance of the summarized output.
Quality of summary
The quality assessment starts with confirming that the correct documents are picked for comparison. Verisk enhanced the quality of the solution by using document metadata to narrow the search results by specifying which documents to include or exclude from a query, resulting in more relevant responses generated by the FM. For the generative AI description of change, Verisk wanted to capture the essence of the change instead of merely highlighting the differences. The results were reviewed by their in-house policy authoring experts and their feedback was used to determine the prompts, document splitting strategy, and FM. With techniques in place to enhance output quality and relevance, Verisk also prioritized optimizing the performance and cost-efficiency of their generative AI solution. These techniques were specific to prompt engineering; some examples are few-shot prompting, chain of thought prompting, and the needle in a haystack approach.
Price-performance
To achieve lower cost, Verisk regularly evaluated various FM options and changed them as new options with lower cost and better performance were released. During the development process, Verisk redesigned the solution to reduce the number of calls to the FM and wherever possible used non-FM based options.
As mentioned earlier, the overall solution consists of a few different components:

Location of the change
Excerpts of the changes
Change summary
Changes shown in the tracked change format

Verisk reduced the FM load and improved accuracy by identifying the sections that contained differences and then passing these sections to the FM to generate the change summary. For constructing the tracked difference format, containing redlines, Verisk used a non-FM based solution. In addition to optimizing performance and cost, Verisk also focused on developing a modular, reusable architecture for their generative AI solution.
Reusability
Good software development practices apply to the development of generative AI solutions too. You can create a decoupled architecture with reusable components. The Mozart generative AI companion is provided as an API, which decouples it from the frontend development and allows for reusability of this capability. Similarly, the API consists of many reusable components like common prompts, common definitions, retrieval service, embedding creation, and persistence service. Through their modular, reusable design approach and iterative optimization process, Verisk was able to achieve highly satisfactory results with their generative AI solution.
Results
Based on Verisk’s evaluation template questions and rounds of testing, they concluded that the results generated over 90% good or acceptable summaries. Testing was done by providing results of the solution to business experts, and having these experts grade the results using a grading scale.
Business impact
Verisk’s customers spend significant time regularly to review changes to the policy forms. The generative AI-powered Mozart companion can simplify the review process by ingesting these complex and unstructured policy documents and providing a summary of changes in minutes. This enables Verisk’s customers to cut the change adoption time from days to minutes. The improved adoption speed not only increases productivity, but also enable timely implementation of changes.
Conclusion
Verisk’s generative AI-powered Mozart companion uses advanced natural language processing and prompt engineering techniques to provide rapid and accurate summaries of changes between insurance policy documents. By harnessing the power of large language models like Anthropic’s Claude 3 Sonnet while incorporating domain expertise, Verisk has developed a solution that significantly accelerates the policy review process for their customers, reducing change adoption time from days or weeks to just minutes. This innovative application of generative AI delivers tangible productivity gains and operational efficiencies to the insurance industry. With a strong governance framework promoting responsible AI use, Verisk is at the forefront of unlocking generative AI’s potential to transform workflows and drive resilience across the global risk landscape.
For more information, see the following resources:

Explore generative AI on AWS
Learn about unlocking the business value of generative AI
Learn more about Anthropic’s Claude 3 models on Amazon Bedrock
Learn about Amazon Bedrock and how to build and scale generative AI applications with FMs
Explore other use cases for generative AI with Amazon Bedrock

About the Authors
Sundeep Sardana is the Vice President of Software Engineering at Verisk Analytics, based in New Jersey. He leads the Reimagine program for the company’s Rating business, driving modernization across core services such as forms, rules, and loss costs. A dynamic change-maker and technologist, Sundeep specializes in building high-performing teams, fostering a culture of innovation, and leveraging emerging technologies to deliver scalable, enterprise-grade solutions. His expertise spans cloud computing, Generative AI, software architecture, and agile development, ensuring organizations stay ahead in an evolving digital landscape. Connect with him on LinkedIn.
Malolan Raman is a Principal Engineer at Verisk, based out of New Jersey specializing in the development of Generative AI (GenAI) applications. With extensive experience in cloud computing and artificial intelligence, He has been at the forefront of integrating cutting-edge AI technologies into scalable, secure, and efficient cloud solutions.
Joseph Lam is the senior director of commercial multi-lines that include general liability, umbrella/excess, commercial property, businessowners, capital assets, crime and inland marine. He leads a team responsible for research, development, and support of commercial casualty products, which mostly consist of forms and rules. The team is also tasked with supporting new and innovative solutions for the emerging marketplace.
Maitri Shah is a Software Development Engineer at Verisk with over two years of experience specializing in developing innovative solutions in Generative AI (GenAI) on Amazon Web Services (AWS). With a strong foundation in machine learning, cloud computing, and software engineering, Maitri has successfully implemented scalable AI models that drive business value and enhance user experiences.
Vaibhav Singh is a Product Innovation Analyst at Verisk, based out of New Jersey. With a background in Data Science, engineering, and management, he works as a pivotal liaison between technology and business, enabling both sides to build transformative products & solutions that tackle some of the current most significant challenges in the insurance domain. He is driven by his passion for leveraging data and technology to build innovative products that not only address the current obstacles but also pave the way for future advancements in that domain.
Ryan Doty is a Solutions Architect Manager at AWS, based out of New York. He helps financial services customers accelerate their adoption of the AWS Cloud by providing architectural guidelines to design innovative and scalable solutions. Coming from a software development and sales engineering background, the possibilities that the cloud can bring to the world excite him.
Tarik Makota is a Sr. Principal Solutions Architect with Amazon Web Services. He provides technical guidance, design advice, and thought leadership to AWS’ customers across the US Northeast. He holds an M.S. in Software Development and Management from Rochester Institute of Technology.
Alex Oppenheim is a Senior Sales Leader at Amazon Web Services, supporting consulting and services customers. With extensive experience in the cloud and technology industry, Alex is passionate about helping enterprises unlock the power of AWS to drive innovation and digital transformation.

Today, Amazon Web Services (AWS) announced the general availability of Amazon Bedrock Knowledge Bases GraphRAG (GraphRAG), a capability in Amazon Bedrock Knowledge Bases that enhances Retrieval-Augmented Generation (RAG) with graph data in Amazon Neptune Analytics. This capability enhances responses from generative AI applications by automatically creating embeddings for semantic search and generating a graph of the entities and relationships extracted from ingested documents. The graph, stored in Amazon Neptune Analytics, provides enriched context during the retrieval phase to deliver more comprehensive, relevant, and explainable responses tailored to customer needs. Developers can enable GraphRAG with just a few clicks on the Amazon Bedrock console to boost the accuracy of generative AI applications without any graph modeling expertise.
In this post, we discuss the benefits of GraphRAG and how to get started with it in Amazon Bedrock Knowledge Bases.
Enhance RAG with graphs for more comprehensive and explainable GenAI applications
Generative AI is transforming how humans interact with technology by having natural conversations that provide helpful, nuanced, and insightful responses. However, a key challenge facing current generative AI systems is providing responses that are comprehensive, relevant, and explainable because data is stored across multiple documents. Without effectively mapping shared context across input data sources, responses risk being incomplete and inaccurate.
To address this, AWS announced a public preview of GraphRAG at re:Invent 2024, and is now announcing its general availability. This new capability integrates the power of graph data modeling with advanced natural language processing (NLP). GraphRAG automatically creates graphs which capture connections between related entities and sections across documents. More specifically, the graph created will connect chunks to documents, and entities to chunks.
During response generation, GraphRAG first does semantic search to find the top k most relevant chunks, and then traverses the surrounding neighborhood of those chunks to retrieve the most relevant content. By linking this contextual information, the generative AI system can provide responses that are more complete, precise, and grounded in source data. Whether answering complex questions across topics or summarizing key details from lengthy reports, GraphRAG delivers the comprehensive and explainable responses needed to enable more helpful, reliable AI conversations.
GraphRAG boosts relevance and accuracy when relevant information is dispersed across multiple sources or documents, which can be seen in the following three use cases.
Streamlining market research to accelerate business decisions
A leading global financial institution sought to enhance insight extraction from its proprietary research. With a vast repository of economic and market research reports, the institution wanted to explore how GraphRAG could improve information retrieval and reasoning for complex financial queries. To evaluate this, they added their proprietary research papers, focusing on critical market trends and economic forecasts.
To evaluate the effectiveness of GraphRAG, the institution partnered with AWS to build a proof-of-concept using Amazon Bedrock Knowledge Bases and Amazon Neptune Analytics. The goal was to determine if GraphRAG could more effectively surface insights compared to traditional retrieval methods. GraphRAG structures knowledge into interconnected entities and relationships, enabling multi-hop reasoning across documents. This capability is crucial for answering intricate questions such as “What are some headwinds and tailwinds to capex growth in the next few years?” or “What is the impact of the ILA strike on international trade?”. Rather than relying solely on keyword matching, GraphRAG allows the model to trace relationships between economic indicators, policy changes, and industry impacts, ensuring responses are contextually rich and data-driven.
When comparing the quality of responses from GraphRAG and other retrieval methods, notable differences emerged in their comprehensiveness, clarity, and relevance. While other retrieval methods delivered straightforward responses, they often lacked deeper insights and broader context. GraphRAG instead provided more nuanced answers by incorporating related factors and offering additional relevant information, which made the responses more comprehensive than the other retrieval methods.
Improving data-driven decision-making in automotive manufacturing
An international auto company manages a large dataset, supporting thousands of use cases across engineering, manufacturing, and customer service. With thousands of users querying different datasets daily, making sure insights are accurate and connected across sources has been a persistent challenge.
To address this, the company worked with AWS to prototype a graph that maps relationships between key data points, such as vehicle performance, supply chain logistics, and customer feedback. This structure allows for more precise results across datasets, rather than relying on disconnected query results.
With Amazon Bedrock Knowledge Bases GraphRAG with Amazon Neptune Analytics automatically constructing a graph from ingested documents, the company can surface relevant insights more efficiently in their RAG applications. This approach helps teams identify patterns in manufacturing quality, predict maintenance needs, and improve supply chain resilience, making data analysis more effective and scalable across the organization.
Enhancing cybersecurity incident analysis
A cybersecurity company is using GraphRAG to improve how its AI-powered assistant analyzes security incidents. Traditional detection methods rely on isolated alerts, often missing the broader context of an attack.
By using a graph, the company connects disparate security signals, such as login anomalies, malware signatures, and network traffic patterns, into a structured representation of threat activity. This allows for faster root cause analysis and more comprehensive security reporting.
Amazon Bedrock Knowledge Bases and Neptune Analytics enable this system to scale while maintaining strict security controls, providing resource isolation. With this approach, the company’s security teams can quickly interpret threats, prioritize responses, and reduce false positives, leading to more efficient incident handling.
Solution overview
In this post, we provide a walkthrough to build Amazon Bedrock Knowledge Bases GraphRAG with Amazon Neptune Analytics, using files in an Amazon Simple Storage Service (Amazon S3) bucket. Running this example will incur costs in Amazon Neptune Analytics, Amazon S3, and Amazon Bedrock. Amazon Neptune Analytics costs for this example will be approximately $0.48 per hour. Amazon S3 costs will vary depending on how large your dataset is, and more details on Amazon S3 pricing can be found here. Amazon Bedrock costs will vary depending on the embeddings model and chunking strategy you select, and more details on Bedrock pricing can be found here.
Prerequisites
To follow along with this post, you need an AWS account with the necessary permissions to access Amazon Bedrock, and an Amazon S3 bucket containing data to serve as your knowledge base. Also ensure that you have enabled model access to Claude 3 Haiku (anthropic.claude-3-haiku-20240307-v1:0) and any other models that you wish to use as your embeddings model. For more details on how to enable model access, refer to the documentation here.
Build Amazon Bedrock Knowledge Bases GraphRAG with Amazon Neptune Analytics
To get started, complete the following steps:

On the Amazon Bedrock console, choose Knowledge Bases under Builder tools in the navigation pane.
In the Knowledge Bases section, choose Create and Knowledge Base with vector store.
For Knowledge Base details, enter a name and an optional description.
For IAM permissions, select Create and use a new service role to create a new AWS Identity and Access Management (IAM) role.
For Data source details, select Amazon S3 as your data source.
Choose Next.
For S3 URI, choose Browse S3 and choose the appropriate S3 bucket.
For Parsing strategy, select Amazon Bedrock default parser.
For Chunking strategy, choose Default chunking (recommended for GraphRAG) or any other strategy as you wish.
Choose Next.
For Embeddings model, choose an embeddings model, such as Amazon Titan Text Embeddings v2.
For Vector database, select Quick create a new vector store and then select Amazon Neptune Analytics (GraphRAG).
Choose Next.
Review the configuration details and choose Create Knowledge Base.

Sync the data source

Once the knowledge base is created, click Sync under the Data source section. The data sync can take a few minutes to a few hours, depending on how many source documents you have and how big each one is.

Test the knowledge base
Once the data sync is complete:

Choose the expansion icon to expand the full view of the testing area.
Configure your knowledge base by adding filters or guardrails.
We encourage you to enable reranking (For information about pricing for reranking models, see Amazon Bedrock Pricing) to fully take advantage of the capabilities of GraphRAG. Reranking allows GraphRAG to refine and optimize search results.
You can also supply a custom metadata file (each up to 10 KB) for each document in the knowledge base. You can apply filters to your retrievals, instructing the vector store to pre-filter based on document metadata and then search for relevant documents. This way, you have control over the retrieved documents, especially if your queries are ambiguous. Note that the list type is not supported.
Use the chat area in the right pane to ask questions about the documents from your Amazon S3 bucket.

The responses will use GraphRAG and provide references to chunks and documents in their response.

Now that you’ve enabled GraphRAG, test it out by querying your generative AI application and observe how the responses have improved compared to baseline RAG approaches. You can monitor the Amazon CloudWatch logs for performance metrics on indexing, query latency, and accuracy.
Clean up
When you’re done exploring the solution, make sure to clean up by deleting any resources you created. Resources to clean up include the Amazon Bedrock knowledge base, the associated AWS IAM role that the Amazon Bedrock knowledge base uses, and the Amazon S3 bucket that was used for the source documents.
You will also need to separately delete the Amazon Neptune Analytics graph that was created on your behalf, by Amazon Bedrock Knowledge Bases.
Conclusion
In this post, we discussed how to get started with Amazon Bedrock Knowledge Bases GraphRAG with Amazon Neptune. For further experimentation, check out the Amazon Bedrock Knowledge Bases Retrieval APIs to use the power of GraphRAG in your own applications. Refer to our documentation for code samples and best practices.

About the authors
Denise Gosnell is a Principal Product Manager for Amazon Neptune, focusing on generative AI infrastructure and graph data applications that enable scalable, cutting-edge solutions across industry verticals.
Melissa Kwok is a Senior Neptune Specialist Solutions Architect at AWS, where she helps customers of all sizes and verticals build cloud solutions according to best practices. When she’s not at her desk you can find her in the kitchen experimenting with new recipes or reading a cookbook.
Ozan Eken is a Product Manager at AWS, passionate about building cutting-edge Generative AI and Graph Analytics products. With a focus on simplifying complex data challenges, Ozan helps customers unlock deeper insights and accelerate innovation. Outside of work, he enjoys trying new foods, exploring different countries, and watching soccer.
Harsh Singh is a Principal Product Manager Technical at AWS AI. Harsh enjoys building products that bring AI to software developers and everyday users to improve their productivity.
Mani Khanuja is a Tech Lead – Generative AI Specialists, author of the book Applied Machine Learning and High-Performance Computing on AWS, and a member of the Board of Directors for Women in Manufacturing Education Foundation Board. She leads machine learning projects in various domains such as computer vision, natural language processing, and generative AI. She speaks at internal and external conferences such AWS re:Invent, Women in Manufacturing West, YouTube webinars, and GHC 23. In her free time, she likes to go for long runs along the beach.

In the rapidly evolving landscape of artificial intelligence, Retrieval Augmented Generation (RAG) has emerged as a game-changer, revolutionizing how Foundation Models (FMs) interact with organization-specific data. As businesses increasingly rely on AI-powered solutions, the need for accurate, context-aware, and tailored responses has never been more critical.
Enter the powerful trio of Amazon Bedrock, LlamaIndex, and RAGAS– a cutting-edge combination that’s set to redefine the evaluation and optimization of RAG responses. This blog post delves into how these innovative tools synergize to elevate the performance of your AI applications, ensuring they not only meet but exceed the exacting standards of enterprise-level deployments.
Whether you’re a seasoned AI practitioner or a business leader exploring the potential of generative AI, this guide will equip you with the knowledge and tools to:

Harness the full potential of Amazon Bedrock robust foundation models
Utilize RAGAS’s comprehensive evaluation metrics for RAG systems

In this post, we’ll explore how to leverage Amazon Bedrock, LlamaIndex, and RAGAS to enhance your RAG implementations. You’ll learn practical techniques to evaluate and optimize your AI systems, enabling more accurate, context-aware responses that align with your organization’s specific needs. Let’s dive in and discover how these powerful tools can help you build more effective and reliable AI-powered solutions.
RAG Evaluation
RAG evaluation is important to ensure that RAG models produce accurate, coherent, and relevant responses. By analyzing the retrieval and generator components both jointly and independently, RAG evaluation helps identify bottlenecks, monitor performance, and improve the overall system. Current RAG pipelines frequently employ similarity-based metrics such as ROUGE, BLEU, and BERTScore to assess the quality of the generated responses, which is essential for refining and enhancing the model’s capabilities.
Above mentioned probabilistic metrics ROUGE, BLEU, and BERTScore have limitations in assessing relevance and detecting hallucinations. More sophisticated metrics are needed to evaluate factual alignment and accuracy.
Evaluate RAG components with Foundation models
We can also use a Foundation Model as a judge to compute various metrics for both retrieval and generation. Here are some examples of these metrics:

Retrieval component

Context precision – Evaluates whether all of the ground-truth relevant items present in the contexts are ranked higher or not.
Context recall – Ensures that the context contains all relevant information needed to answer the question.

Generator component

Faithfulness – Verifies that the generated answer is factually accurate based on the provided context, helping to identify errors or “hallucinations.”
Answer relavancy : Measures how well the answer matches the question. Higher scores mean the answer is complete and relevant, while lower scores indicate missing or redundant information.

Overview of solution
This post guides you through the process of assessing quality of RAG response with evaluation framework such as RAGAS and LlamaIndex with Amazon Bedrock.
In this post, we are also going to leverage Langchain to create a sample RAG application.
Amazon Bedrock is a fully managed service that offers a choice of high-performing Foundation Models (FMs) from leading AI companies like AI21 Labs, Anthropic, Cohere, Meta, Stability AI, and Amazon via a single API, along with a broad set of capabilities you need to build generative AI applications with security, privacy, and responsible AI.
The Retrieval Augmented Generation Assessment (RAGAS) framework offers multiple metrics to evaluate each part of the RAG system pipeline, identifying areas for improvement. It utilizes foundation models to test individual components, aiding in pinpointing modules for development to enhance overall results.
LlamaIndex is a framework for building LLM applications. It simplifies data integration from various sources and provides tools for data indexing, engines, agents, and application integrations. Optimized for search and retrieval, it streamlines querying LLMs and retrieving documents. This blog post focuses on using its Observability/Evaluation modules.
LangChain is an open-source framework that simplifies the creation of applications powered by foundation models. It provides tools for chaining LLM operations, managing context, and integrating external data sources. LangChain is primarily used for building chatbots, question-answering systems, and other AI-driven applications that require complex language processing capabilities.
Diagram Architecture
The following diagram is a high-level reference architecture that explains how you can evaluate the RAG solution with RAGAS or LlamaIndex.

The solution consists of the following components:

Evaluation dataset – The source data for the RAG comes from the Amazon SageMaker FAQ, which represents 170 question-answer pairs. This corresponds to Step 1 in the architecture diagram.

Build sample RAG – Documents are segmented into chunks and stored in an Amazon Bedrock Knowledge Bases (Steps 2–4). We use Langchain Retrieval Q&A to answer user queries. This process retrieves relevant data from an index at runtime and passes it to the Foundation Model (FM).
RAG evaluation – To assess the quality of the Retrieval-Augmented Generation (RAG) solution, we can use both RAGAS and LlamaIndex. An LLM performs the evaluation by comparing its predictions with ground truths (Steps 5–6).

You must follow the provided notebook to reproduce the solution. We elaborate on the main code components in this post.
Prerequisites
To implement this solution, you need the following:

An AWS accountwith privileges to create AWS Identity and Access Management (IAM) roles and policies. For more information, see Overview of access management: Permissions and policies.
Access enabled for the Amazon Titan Embeddings G1 – Text model and Anthropic Claude 3 Sonnet on Amazon Bedrock. For instructions, see Model access.
Run the prerequisite code provided in the Python

Ingest FAQ data
The first step is to ingest the SageMaker FAQ data. For this purpose, LangChain provides a WebBaseLoader object to load text from HTML webpages into a document format. Then we split each document in multiple chunks of 2,000 tokens with a 100-token overlap. See the following code below:
text_chunks = split_document_from_url(SAGEMAKER_URL, chunck_size= 2000,  chunk_overlap=100)
retriever_db= get_retriever(text_chunks, bedrock_embeddings)
Set up embeddings and LLM with Amazon Bedrock and LangChain
In order to build a sample RAG application, we need an LLM and an embedding model:

LLM – Anthropic Claude 3 Sonnet

Embedding – Amazon Titan Embeddings – Text V2

This code sets up a LangChain application using Amazon Bedrock, configuring embeddings with Titan and a Claude 3 Sonnet model for text generation with specific parameters for controlling the model’s output. See the following code below from the notebook :
from botocore.client import Config
from langchain.llms.bedrock import Bedrock
from langchain_aws import ChatBedrock
from langchain.embeddings import BedrockEmbeddings
from langchain.retrievers.bedrock import AmazonKnowledgeBasesRetriever
from langchain.chains import RetrievalQA
import nest_asyncio
nest_asyncio.apply()

#URL to fetch the document
SAGEMAKER_URL=”https://aws.amazon.com/sagemaker/faqs/”

#Bedrock parameters
EMBEDDING_MODEL=”amazon.titan-embed-text-v2:0″
BEDROCK_MODEL_ID=”anthropic.claude-3-sonnet-20240229-v1:0″

bedrock_embeddings = BedrockEmbeddings(model_id=EMBEDDING_MODEL,client=bedrock_client)

model_kwargs = {
“temperature”: 0,
“top_k”: 250,
“top_p”: 1,
“stop_sequences”: [“\n\nHuman:”]
}

llm_bedrock = ChatBedrock(
model_id=BEDROCK_MODEL_ID,
model_kwargs=model_kwargs
)
Set up Knowledge Bases
We will create Amazon Bedrock knowledgebases Web Crawler datasource and process Sagemaker FAQ data.
In the code below, we load the embedded documents in Knowledge bases and we set up the retriever with LangChain:
from utils import split_document_from_url, get_bedrock_retriever
from botocore.exceptions import ClientError

text_chunks = split_document_from_url(SAGEMAKER_URL, chunck_size= 2000,  chunk_overlap=100)
retriever_db= get_bedrock_retriever(text_chunks, region)
Build a Q&A chain to query the retrieval API
After the database is populated, create a Q&A retrieval chain to perform question answering with context extracted from the vector store. You also define a prompt template following Claude prompt engineering guidelines. See the following code below from the notebook:
from langchain_core.prompts import ChatPromptTemplate
from langchain.chains.combine_documents import create_stuff_documents_chain
from langchain.chains import create_retrieval_chain

system_prompt = (
“Use the given context to answer the question. ”
“If you don’t know the answer, say you don’t know. ”
“Use three sentence maximum and keep the answer concise and short. ”
“Context: {context}”
)

prompt_template = ChatPromptTemplate.from_messages([
(“system”, system_prompt),
(“human”, “{input}”)
]
)
question_answer_chain = create_stuff_documents_chain(llm_bedrock, prompt_template)
chain = create_retrieval_chain(retriever_db, question_answer_chain)
Build Dataset to evaluate RAG application
To evaluate a RAG application, we need a combination of the following datasets:

Questions – The user query that serves as input to the RAG pipeline
Context – The information retrieved from enterprise or external data sources based on the provided query
Answers – The responses generated by LLMs
Ground truths – Human-annotated, ideal responses for the questions that can be used as the benchmark to compare against the LLM-generated answers

We are ready to evaluate the RAG application. As describe in the introduction, we select 3 metrics to assess our RAG solution:

Faithfulness
Answer Relevancy
Answer Correctness

For more information, refer to Metrics.
This step involves defining an evaluation dataset with a set of ground truth questions and answers. For this post, we choose four random questions from the SageMaker FAQ. See the following code below from the notebook:
EVAL_QUESTIONS = [
“Can I stop a SageMaker Autopilot job manually?”,
“Do I get charged separately for each notebook created and run in SageMaker Studio?”,
“Do I get charged for creating and setting up an SageMaker Studio domain?”,
“Will my data be used or shared to update the base model that is offered to customers using SageMaker JumpStart?”,
]

#Defining the ground truth answers for each question

EVAL_ANSWERS = [
“Yes. You can stop a job at any time. When a SageMaker Autopilot job is stopped, all ongoing trials will be stopped and no new trial will be started.”,
“””No. You can create and run multiple notebooks on the same compute instance.
You pay only for the compute that you use, not for individual items.
You can read more about this in our metering guide.
In addition to the notebooks, you can also start and run terminals and interactive shells in SageMaker Studio, all on the same compute instance.”””,
“No, you don’t get charged for creating or configuring an SageMaker Studio domain, including adding, updating, and deleting user profiles.”,
“No. Your inference and training data will not be used nor shared to update or train the base model that SageMaker JumpStart surfaces to customers.”
]
Evaluation of RAG with RAGAS
Evaluating the RAG solution requires to compare LLM predictions with ground truth answers. To do so, we use the batch() function from LangChain to perform inference on all questions inside our evaluation dataset.
Then we can use the evaluate() function from RAGAS to perform evaluation on each metric (answer relevancy, faithfulness and answer corectness). It uses an LLM to compute metrics. Feel free to use other Metrics from RAGAS.
See the following code below from the notebook:
from ragas.metrics import answer_relevancy, faithfulness, answer_correctness
from ragas import evaluate

#Batch invoke and dataset creation
result_batch_questions = chain.batch([{“input”: q} for q in EVAL_QUESTIONS])

dataset= build_dataset(EVAL_QUESTIONS,EVAL_ANSWERS,result_batch_questions, text_chunks)

result = evaluate(dataset=dataset, metrics=[ answer_relevancy, faithfulness, answer_correctness ],llm=llm_bedrock, embeddings=bedrock_embeddings, raise_exceptions=False )
df = result.to_pandas()
df.head()
The following screenshot shows the evaluation results and the RAGAS answer relevancy score.

Answer Relevancy
In the answer_relevancy_score column, a score closer to 1 indicates the response generated is relevant to the input query.
Faithfulness
In the second column, the first query result has a lower faithfulness_score (0.2), which indicates the responses are not derived from the context and are hallucinations. The rest of the query results have a higher faithfulness_score (1.0), which indicates the responses are derived from the context.
Answer Correctness
In the last column answer_correctness, the second and last row have high answer correctness, meaning that answer provided by the LLM is closer to to from the groundtruth.
Evaluation of RAG with LlamaIndex
LlamaIndex, similar to Ragas, provides a comprehensive RAG (Retrieval-Augmented Generation) evaluation module. This module offers a variety of metrics to assess the performance of your RAG system. The evaluation process generates two key outputs:

Feedback: The judge LLM (Language Model) provides detailed evaluation feedback in the form of a string, offering qualitative insights into the system’s performance.
Score: This numerical value indicates how well the answer meets the evaluation criteria. The scoring system varies depending on the specific metric being evaluated. For example, metrics like Answer Relevancy and Faithfulness are typically scored on a scale from 0 to 1.

These outputs allow for both qualitative and quantitative assessment of your RAG system’s performance, enabling you to identify areas for improvement and track progress over time.
The following is a code sample from the notebook:
from llama_index.llms.bedrock import Bedrock
from llama_index.core.evaluation import (
AnswerRelevancyEvaluator,
CorrectnessEvaluator,
FaithfulnessEvaluator
)
from utils import evaluate_llama_index_metric

bedrock_llm_llama = Bedrock(model=BEDROCK_MODEL_ID)
faithfulness= FaithfulnessEvaluator(llm=bedrock_llm_llama)
answer_relevancy= AnswerRelevancyEvaluator(llm=bedrock_llm_llama)
correctness= CorrectnessEvaluator(llm=bedrock_llm_llama)
Answer Relevancy
df_answer_relevancy= evaluate_llama_index_metric(answer_relevancy, dataset)
df_answer_relevancy.head()

The column Score defines the result for the answer_relevancy evaluation criteria. All passing values are set to 1, meaning that all predictions are relevant with the context retrieved.
Additionally, the column Feedback provides a clear explanation of the result of the passing score. We can observe that all answers align with the context extracted from the retriever.
Answer Correctness
df_correctness= evaluate_llama_index_metric(correctness, dataset)
df_correctness.head()

All values from the column Score are set to 5.0, meaning that all predictions are coherent with ground truth answers.
Faithfulness
The following screenshot shows the evaluation results for answer faithfulness.
df_faithfulness= evaluate_llama_index_metric(faithfulness, dataset)
df_faithfulness.head()

All values from the Score column are set to 1.0, which means all answers generated by LLM are coherent given the context retrieved.
Conclusion
While Foundation Models offer impressive generative capabilities, their effectiveness in addressing organization-specific queries has been a persistent challenge. The Retrieval Augmented Generation framework emerges as a powerful solution, bridging this gap by enabling LLMs to leverage external, organization-specific data sources.
To truly unlock the potential of RAG pipelines, the RAGAS framework, in conjunction with LlamaIndex, provides a comprehensive evaluation solution. By meticulously assessing both retrieval and generation components, this approach empowers organizations to pinpoint areas for improvement and refine their RAG implementations. The result? Responses that are not only factually accurate but also highly relevant to user queries.
By adopting this holistic evaluation approach, enterprises can fully harness the transformative power of generative AI applications. This not only maximizes the value derived from these technologies but also paves the way for more intelligent, context-aware, and reliable AI systems that can truly understand and address an organization’s unique needs.
As we continue to push the boundaries of what’s possible with AI, tools like Amazon Bedrock, LlamaIndex, and RAGAS will play a pivotal role in shaping the future of enterprise AI applications. By embracing these innovations, organizations can confidently navigate the exciting frontier of generative AI, unlocking new levels of efficiency, insight, and competitive advantage.
For further exploration, readers interested in enhancing the reliability of AI-generated content may want to look into Amazon Bedrock’s Guardrails feature, which offers additional tools like the Contextual Grounding Check.

About the authors
Madhu is a Senior Partner Solutions Architect specializing in worldwide public sector cybersecurity partners. With over 20 years in software design and development, he collaborates with AWS partners to ensure customers implement solutions that meet strict compliance and security objectives. His expertise lies in building scalable, highly available, secure, and resilient applications for diverse enterprise needs.
Babu Kariyaden Parambath is a Senior AI/ML Specialist at AWS. At AWS, he enjoys working with customers in helping them identify the right business use case with business value and solve it using AWS AI/ML solutions and services. Prior to joining AWS, Babu was an AI evangelist with 20 years of diverse industry experience delivering AI driven business value for customers.

This post was co-authored with Johann Wildgruber, Dr. Jens Kohl, Thilo Bindel, and Luisa-Sophie Gloger from BMW Group.
The BMW Group—headquartered in Munich, Germany—is a vehicle manufacturer with more than 154,000 employees, and 30 production and assembly facilities worldwide as well as research and development locations across 17 countries. Today, the BMW Group (BMW) is the world’s leading manufacturer of premium automobiles and motorcycles, and provider of premium financial and mobility services.
BMW Connected Company is a division within BMW responsible for developing and operating premium digital services for BMW’s connected fleet, which currently numbers more than 23 million vehicles worldwide. These digital services are used by many BMW vehicle owners daily; for example, to lock or open car doors remotely using an app on their phone, to start window defrost remotely, to buy navigation map updates from the car’s menu, or to listen to music streamed over the internet in their car.
In this post, we explain how BMW uses generative AI technology on AWS to help run these digital services with high availability. Specifically, BMW uses Amazon Bedrock Agents to make remediating (partial) service outages quicker by speeding up the otherwise cumbersome and time-consuming process of root cause analysis (RCA). The fully automated RCA agent correctly identifies the right root cause for most cases (measured at 85%), and helps engineers in terms of system understanding and real-time insights in their cases. This performance was further validated during the proof of concept, where employing the RCA agent on representative use cases clearly demonstrates the benefits of this solution, allowing BMW to achieve significantly lower diagnosis times.
The challenges of root cause analysis
Digital services are often implemented by chaining multiple software components together; components that might be built and run by different teams. For example, consider the service of remotely opening and locking vehicle doors. There might be a development team building and running the iOS app, another team for the Android app, a team building and running the backend-for-frontend used by both the iOS and Android app, and so on. Moreover, these teams might be geographically dispersed and run their workloads in different locations and regions; many hosted on AWS, some elsewhere.
Now consider a (fictitious) scenario where reports come in from car owners complaining that remotely locking doors with the app no longer works. Is the iOS app responsible for the outage, or the backend-for-frontend? Did a firewall rule change somewhere? Did an internal TLS certificate expire? Is the MQTT system experiencing delays? Was there an inadvertent breaking change in recent API changes? When did they actually deploy that? Or was the database password for the central subscription service rotated again?
It can be difficult to determine the root cause of issues in situations like this. It requires checking many systems and teams, many of which might be failing, because they’re interdependent. Developers need to reason about the system architecture, form hypotheses, and follow the chain of components until they have located the one that is the culprit. They often have to backtrack and reassess their hypotheses, and pursue the investigation in another chain of components.
Understanding the challenges in such complex systems highlights the need for a robust and efficient approach to root cause analysis. With this context in mind, let’s explore how BMW and AWS collaborated to develop a solution using Amazon Bedrock Agents to streamline and enhance the RCA process.
Solution overview
At a high level, the solution uses an Amazon Bedrock agent to do automated RCA. This agent has several custom-built tools at its disposal to do its job. These tools, implemented by AWS Lambda functions, use services like Amazon CloudWatch and AWS CloudTrail to analyze system logs and metrics. The following diagram illustrates the solution architecture.

When an incident occurs, an on-call engineer gives a description of the issue at hand to the Amazon Bedrock agent. The agent will then start investigating for the root cause of the issue, using its tools to do tasks that the on-call engineer would otherwise do manually, such as searching through logs. Based on the clues it uncovers, the agent proposes several likely hypotheses to the on-call engineer. The engineer can then resolve the issue, or give pointers to the agent to direct the investigation further. In the following section, we take a closer look at the tools the agent uses.
Amazon Bedrock agent tools
The Amazon Bedrock agent’s effectiveness in performing RCA lies in its ability to seamlessly integrate with custom tools. These tools, designed as Lambda functions, use AWS services like CloudWatch and CloudTrail to automate tasks that are typically manual and time-intensive for engineers. By organizing its capabilities into specialized tools, the Amazon Bedrock agent makes sure that RCA is both efficient and precise.
Architecture Tool
The Architecture Tool uses C4 diagrams to provide a comprehensive view of the system’s architecture. These diagrams, enhanced through Structurizr, give the agent a hierarchical understanding of component relationships, dependencies, and workflows. This allows the agent to target the most relevant areas during its RCA process, effectively narrowing down potential causes of failure based on how different systems interact.
For instance, if an issue affects a specific service, the Architecture Tool can identify upstream or downstream dependencies and suggest hypotheses focused on those systems. This accelerates diagnostics by enabling the agent to reason contextually about the architecture instead of blindly searching through logs or metrics.
Logs Tool
The Logs Tool uses CloudWatch Logs Insights to analyze log data in real time. By searching for patterns, errors, or anomalies, as well as comparing the trend to the previous period, it helps the agent pinpoint issues related to specific events, such as failed authentications or system crashes.
For example, in a scenario involving database access failures, the Logs Tool might identify a new spike in the number of error messages such as “FATAL: password authentication failed” compared to the previous hour. This insight allows the agent to quickly associate the failure with potential root causes, such as an improperly rotated database password.
Metrics Tool
The Metrics Tool provides the agent with real-time insights into the system’s health by monitoring key metrics through CloudWatch. This tool identifies statistical anomalies in critical performance indicators such as latency, error rates, resource utilization, or unusual spikes in usage patterns, which can often signal potential issues or deviations from normal behavior.
For instance, in a Kubernetes memory overload scenario, the Metrics Tool might detect a sharp increase in memory consumption or unusual resource allocation prior to the failure. By surfacing CloudWatch metric alarms for such anomalies, the tool enables the agent to prioritize hypotheses related to resource mismanagement, misconfigured thresholds, or unexpected system load, guiding the investigation more effectively toward resolving the issue.
Infrastructure Tool
The Infrastructure Tool uses CloudTrail data to analyze critical control-plane events, such as configuration changes, security group updates, or API calls. This tool is particularly effective in identifying misconfigurations or breaking changes that might trigger cascading failures.
Consider a case where a security group ingress rule is inadvertently removed, causing connectivity issues between services. The Infrastructure Tool can detect and correlate this event with the reported incident, providing the agent with actionable insights to guide its RCA process.
By combining these tools, the Amazon Bedrock agent mimics the step-by-step reasoning of an experienced engineer while executing tasks at machine speed. The modular nature of the tools allows for flexibility and customization, making sure that RCA is tailored to the unique needs of BMW’s complex, multi-regional cloud infrastructure.
In the next section, we discuss how these tools work together within the agent’s workflow.
Amazon Bedrock agents: The ReAct framework in action
At the heart of BMW’s rapid RCA lies the ReAct (Reasoning and Action) agent framework, an innovative approach that dynamically combines logical reasoning with task execution. By integrating ReAct with Amazon Bedrock, BMW gains a flexible solution for diagnosing and resolving complex cloud-based incidents. Unlike traditional methods, which rely on predefined workflows, ReAct agents use real-time inputs and iterative decision-making to adapt to the specific circumstances of an incident.
The ReAct agent in BMW’s RCA solution uses a structured yet adaptive workflow to diagnose and resolve issues. First, it interprets the textual description of an incident (for example, “Vehicle doors cannot be locked via the app”) to identify which parts of the system are most likely impacted. Guided by the ReAct framework’s iterative reasoning, the agent then gathers evidence by calling specialized tools, using data centrally aggregated in a cross-account observability setup. By continuously reevaluating the results of each tool invocation, the agent zeros in on potential causes—whether an expired certificate, a revoked firewall rule, or a spike in traffic—until it isolates the root cause. The following diagram illustrates this workflow.

The ReAct framework offers the following benefits:

Dynamic and adaptive – The ReAct agent tailors its approach to the specific incident, rather than a one-size-fits-all methodology. This adaptability is especially critical in BMW’s multi-regional, multi-service architecture.
Efficient tool utilization – By reasoning about which tools to invoke and when, the ReAct agent minimizes redundant queries, providing faster diagnostics without overloading AWS services like CloudWatch or CloudTrail.
Human-like reasoning – The ReAct agent mimics the logical thought process of a seasoned engineer, iteratively exploring hypotheses until it identifies the root cause. This capability bridges the gap between automation and human expertise.

By employing Amazon Bedrock ReAct agents, significantly lower diagnosis times are achieved. These agents not only enhance operational efficiency but also empower engineers to focus on strategic improvements rather than labor-intensive diagnostics.
Case study: Root cause analysis “Unlocking vehicles via the iOS app”
To illustrate the power of Amazon Bedrock agents in action, let us explore a possible real-world scenario involving the interplay between BMW’s connected fleet and the digital services running in the cloud backend.
We deliberately change the security group for the central networking account in a test environment. This has the effect that requests from the fleet are (correctly) blocked by the changed security group and do not reach the services hosted in the backend. Hence, a test user cannot lock or unlock her vehicle door remotely.
Incident details
BMW engineers received a report from a tester indicating the remote lock/unlock functionality on the mobile app does not work.
This report raised immediate questions: was the issue in the app itself, the backend-for-frontend service, or deeper within the system, such as in the MQTT connectivity or authentication mechanisms?
How the ReAct agent addresses the problem
The problem is described to the Amazon Bedrock ReAct agent: “Users of the iOS app cannot unlock car doors remotely.” The agent immediately begins its analysis:

The agent begins by understanding the overall system architecture, calling the Architecture Tool. The outputs of the architecture tool reveal that the iOS app, like the Android app, is connected to a backend-for-frontend API, and that the backend-for-frontend API itself is connected to several other internal APIs, such as the Remote Vehicle Management API. The Remote Vehicle Management API is responsible for sending commands to cars by using MQTT messaging.
The agent uses the other tools at its disposal in a targeted way: it scans the logs, metrics, and control plane activities of only those components that are involved in remotely unlocking car doors: iOS app remote logs, backend-for-frontend API logs, and so on. The agent finds several clues:

Anomalous logs that indicate connectivity issues (network timeouts).
A sharp decrease in the number of successful invocations of the Remote Vehicle Management API.
Control plane activities: several security groups in the central networking account hosted on the testing environment were changed.

Based on those findings, the agent infers and defines several hypotheses and presents these to the user, ordered by their likelihood. In this case, the first hypothesis is the actual root cause: a security group was inadvertently changed in the central networking account, which meant that network traffic between the backend-for-frontend and the Remote Vehicle Management API was now blocked. The agent correctly correlated logs (“fetch timeout error”), metrics (decrease in invocations) and control plane changes (security group ingress rule removed) to come to this conclusion.
If the on-call engineer wants further information, they can now ask follow-up questions to the agent, or instruct the agent to investigate elsewhere as well.

The entire process—from incident detection to resolution—took minutes, compared to the hours it could have taken with traditional RCA methods. The ReAct agent’s ability to dynamically reason, access cross-account observability data, and iterate on its hypotheses alleviated the need for tedious manual investigations.

Conclusion
By using Amazon Bedrock ReAct agents, BMW has shown how to improve its approach to root cause analysis, turning a complex and manual process into an efficient, automated workflow. The tools integrated within the ReAct framework significantly narrow down potential reasoning space, and enable dynamic hypotheses generation and targeted diagnostics, mimicking the reasoning process of seasoned engineers while operating at machine speed. This innovation has reduced the time required to identify and resolve service disruptions, further enhancing the reliability of BMW’s connected services and improving the experience for millions of customers worldwide.
The solution has demonstrated measurable success, with the agent identifying root causes in 85% of test cases and providing detailed insights in the remainder, greatly expediting engineers’ investigations. By lowering the barrier to entry for junior engineers, it has enabled less-experienced team members to diagnose issues effectively, maintaining reliability and scalability across BMW’s operations.
Incorporating generative AI into RCA processes showcases the transformative potential of AI in modern cloud-based operations. The ability to adapt dynamically, reason contextually, and handle complex, multi-regional infrastructures makes Amazon Bedrock Agents a game changer for organizations aiming to maintain high availability in their digital services.
As BMW continues to expand its connected fleet and digital offerings, the adoption of generative AI-driven solutions like Amazon Bedrock will play an important role in maintaining operational excellence and delivering seamless experiences to customers. By following BMW’s example, your organization can also benefit from Amazon Bedrock Agents for root cause analysis to enhance service reliability.
Get started by exploring Amazon Bedrock Agents to optimize your incident diagnostics or use CloudWatch Logs Insights to identify anomalies in your system logs. If you want a hands-on introduction to creating your own Amazon Bedrock agents—complete with code examples and best practices—check out the following GitHub repo. These tools are setting a new industry standard for efficient RCA and operational excellence.

About the Authors
Johann Wildgruber is a transformation lead reliability engineer at BMW Group, working currently to set up an observability platform to strengthen the reliability of ConnectedDrive services. Johann has several years of experience as a product owner in operating and developing large and complex cloud solutions. He is interested in applying new technologies and methods in software development.
Dr. Jens Kohl is a technology leader and builder with over 13 years of experience at the BMW Group. He is responsible for shaping the architecture and continuous optimization of the Connected Vehicle cloud backend. Jens has been leading software development and machine learning teams with a focus on embedded, distributed systems and machine learning for more than 10 years.
Thilo Bindel is leading the Offboard Reliability & Data Engineering team at BMW Group. He is responsible for defining and implementing strategies to ensure reliability, availability, and maintainability of BMW’s backend services in the Connected Vehicle domain. His goal is to establish reliability and data engineering best practices consistently across the organization and to position the BMW Group as a leader in data-driven observability within the automotive industry and beyond.
Luisa-Sophie Gloger is a Data Scientist at the BMW Group with a focus on Machine Learning. As a lead developer within the Connected Company’s Connected AI platform team, she enjoys helping teams to improve their products and workflows with Generative AI. She also has a background in working on Natural Language processing (NLP) and a degree in psychology.
Tanrajbir Takher is a Data Scientist at AWS’s Generative AI Innovation Center, where he works with enterprise customers to implement high-impact generative AI solutions. Prior to AWS, he led research for new products at a computer vision unicorn and founded an early generative AI startup.
Otto Kruse is a Principal Solutions Developer within AWS Industries – Prototyping and Customer Engineering (PACE), a multi-disciplinary team dedicated to helping large companies utilize the potential of the AWS cloud by exploring and implementing innovative ideas. Otto focuses on application development and security.
Huong Vu is a Data Scientist at AWS Generative AI Innovation Centre. She drives projects to deliver generative-AI applications for enterprise customers from a diverse range of industries. Prior to AWS, she worked on improving NLP models for Alexa shopping assistant both on the Amazon.com website and on Echo devices.
Aishwarya is a Senior Customer Solutions Manager with AWS Automotive. She is passionate about solving business problems using Generative AI and cloud-based technologies.
Satyam Saxena is an Applied Science Manager at AWS Generative AI Innovation Center team. He leads Generative AI customer engagements, driving innovative ML/AI initiatives from ideation to production with over a decade of experience in machine learning and data science. His research interests include deep learning, computer vision, NLP, recommender systems, and generative AI.
Kim Robins, a Senior AI Strategist at AWS’s Generative AI Innovation Center, leverages his extensive artificial intelligence and machine learning expertise to help organizations develop innovative products and refine their AI strategies, driving tangible business value.

Generative AI question-answering applications are pushing the boundaries of enterprise productivity. These assistants can be powered by various backend architectures including Retrieval Augmented Generation (RAG), agentic workflows, fine-tuned large language models (LLMs), or a combination of these techniques. However, building and deploying trustworthy AI assistants requires a robust ground truth and evaluation framework.
Ground truth data in AI refers to data that is known to be factual, representing the expected use case outcome for the system being modeled. By providing an expected outcome to measure against, ground truth data unlocks the ability to deterministically evaluate system quality. Running deterministic evaluation of generative AI assistants against use case ground truth data enables the creation of custom benchmarks. These benchmarks are essential for tracking performance drift over time and for statistically comparing multiple assistants in accomplishing the same task. Additionally, they enable quantifying performance changes as a function of enhancements to the underlying assistant, all within a controlled setting. With deterministic evaluation processes such as the Factual Knowledge and QA Accuracy metrics of FMEval, ground truth generation and evaluation metric implementation are tightly coupled. To ensure the highest quality measurement of your question answering application against ground truth, the evaluation metric’s implementation must inform ground truth curation.
In this post, we discuss best practices for applying LLMs to generate ground truth for evaluating question-answering assistants with FMEval on an enterprise scale. FMEval is a comprehensive evaluation suite from Amazon SageMaker Clarify, and provides standardized implementations of metrics to assess quality and responsibility. To learn more about FMEval, see Evaluate large language models for quality and responsibility of LLMs. Additionally, see the Generative AI Security Scoping Matrix for guidance on moderating confidential and personally identifiable information (PII) as part of your generative AI solution.
By following these guidelines, data teams can implement high fidelity ground truth generation for question-answering use case evaluation with FMEval. For ground truth curation best practices for question answering evaluations with FMEval that you can use to design FMEval ground truth prompt templates, see Ground truth curation and metric interpretation best practices for evaluating generative AI question answering using FMEval.
Generating ground truth for FMEval question-answering evaluation
One option to get started with ground truth generation is human curation of a small question-answer dataset. The human curated dataset should be small (based on bandwidth), high in signal, and ideally prepared by use case subject matter experts (SMEs). The exercise of generating this dataset forces a data alignment exercise early in the evaluation process, raising important questions and conversations among use case stakeholders about what questions are important to measure over time for the business. The outcomes for this exercise are three-fold:

Stakeholder alignment on the top N important questions
Stakeholder awareness of the evaluation process
A high-fidelity starter ground truth dataset for the first proof of concept evaluation as a function of awareness and evaluation

While an SME ground truth curation exercise is a strong start, at the scale of an enterprise knowledge base, pure SME generation of ground truth will become prohibitively time and resource intensive. To scale ground truth generation and curation, you can apply a risk-based approach in conjunction with a prompt-based strategy using LLMs. It’s important to note that LLM-generated ground truth isn’t a substitute for use case SME involvement. For example, if ground truth is generated by LLMs before the involvement of SMEs, SMEs will still be needed to identify which questions are fundamental to the business and then align the ground truth with business value as part of a human-in-the-loop process.
To demonstrate, we provide a step-by-step walkthrough using Amazon’s 2023 letter to shareholders as source data.
In keeping with ground truth curation best practices for FMEval question-answering, ground truth is curated as question-answer-fact triplets. The question and answer are curated to suit the ideal question-answering assistant response in terms of content, length, and style. The fact is a minimal representation of the ground truth answer, comprising one or more subject entities of the question.
For example, consider how the following source document chunk from the Amazon 2023 letter to shareholders can be converted to question-answering ground truth.

Dear Shareholders:
Last year at this time, I shared my enthusiasm and optimism for Amazon’s future. Today, I have even more. The reasons are many, but start with the progress we’ve made in our financial results and customer experiences, and extend to our continued innovation and the remarkable opportunities in front of us. In 2023, Amazon’s total revenue grew 12% year-over-year (“Y oY”) from $514B to $575B. By segment, North America revenue increased 12% Y oY from $316B to $353B, International revenue grew 11% Y oY from$118B to $131B, and AWS revenue increased 13% Y oY from $80B to $91B. Further, Amazon’s operating income and Free Cash Flow (“FCF”) dramatically improved. Operating income in 2023 improved 201% YoY from $12.2B (an operating margin of 2.4%) to $36.9B (an operating margin of 6.4%).

To convert the source document excerpt into ground truth, we provide a base LLM prompt template. In the template, we instruct the LLM to take a fact-based approach to interpreting the chunk using chain-of-thought logic. For our example, we work with Anthropic’s Claude LLM on Amazon Bedrock. The template is compatible with and can be modified for other LLMs, such as LLMs hosted on Amazon Sagemaker Jumpstart and self-hosted on AWS infrastructure. To modify the prompt for use by other LLMs, a different approach to denoting prompt sections than XML tags might be required. For example, Meta Llama models apply tags such as <s> [INST] and <<SYS>>. For more information, see the Amazon Bedrock documentation on LLM prompt design and the FMEval documentation.
The LLM is assigned a persona to set its point of view for carrying out the task. In the instructions, the LLM identifies facts as entities from the source document chunk. For each fact, a question-answer-fact triplet is assembled based on the fact detected and its surrounding context. In the prompt, we provide detailed examples for controlling the content of ground truth questions. The examples focus on questions on chunk-wise business knowledge while ignoring irrelevant metadata that might be contained in a chunk. You can customize the prompt examples to fit your ground truth use case.
We further instruct the LLM to apply ground truth curation best practices for FMEval, such as generating multiple variations of facts to fit multiple possible unit expressions. Additional curation elements subject to the task at hand—such as brand language and tone—can be introduced into the ground truth generation prompt. With the following template, we verified that Anthropic’s Claude Sonnet 3.5 can generate custom ground truth attributes accommodating FMEval features, such as the <OR> delimiter to denote alternative acceptable answers for a ground truth fact.

“””You are an expert in ground truth curation for generative AI application evaluation on AWS.

Follow the instructions provided in the <instructions> XML tag for generating question answer fact triplets from a source document excerpt.

<instructions>
– Let’s work this out in a step-by-step way to be sure we have the right answer.
– Review the source document excerpt provided in <document> XML tags below
– For each meaningful domain fact in the <document>, extract an unambiguous question-answer-fact set in JSON format including a question and answer pair encapsulating the fact in the form of a short sentence, followed by a minimally expressed fact extracted from the answer.

<domain_knowledge_focus>
– Focus ONLY on substantive domain knowledge contained within the document content
– Ignore all metadata and structural elements including but not limited to:
– Document dates, versions, page numbers
– Section numbers or titles
– Table structure or row/column positions
– List positions or ordering
– Questions must reference specific domain entities rather than generic document elements
</domain_knowledge_focus>

<context_specification_requirements>
Document Source Identification
– Always reference the specific source document and its date/version
– Example: “According to the [Document Name + Date], what is [specific query]?”

Cross-Reference Prevention
– Each question must be answerable from the current document chunk only
– Do not create questions requiring information from multiple documents
– Example: “In this [Document Name], what are [specific requirements]?”

Department/LOB Specification
– Always specify the relevant department, line of business, or organizational unit
– Example: “What are the [Department Name]’s requirements for [specific process]?”

Document Section Targeting
– Reference specific sections when the information location is relevant
– Example: “In Section [X] of [Document Name], what are the steps for [specific process]?”

Role-Based Context
– Specify relevant roles, responsibilities, or authority levels
– Example: “Which [specific roles] are authorized to [specific action]?”

Version Control Elements
– Include relevant version or revision information
– Example: “What changes were implemented in the [Month Year] revision of [Document]?”

Policy/Procedure Numbers
– Include specific policy or procedure reference numbers
– Example: “Under Policy [Number], what are the requirements for [specific action]?”

Regulatory Framework References
– Specify relevant regulatory frameworks or compliance requirements
– Example: “What [Regulation] compliance requirements are specified for [specific process]?”

System/Platform Specification
– Name specific systems, platforms, or tools
– Example: “What steps are required in [System Name] to [specific action]?”

Document Type Classification
– Specify the type of document (SOP, Policy, Manual, etc.)
– Example: “In the [Document Type + Number], where is [specific information] stored?”

Temporal Validity
– Include effective dates or time periods
– Example: “What process is effective from [Date] according to [Document]?”

Geographic Jurisdiction
– Specify relevant geographic regions or jurisdictions
– Example: “What requirements apply to [Region] according to [Document]?”

Business Process Owner
– Identify relevant process owners or responsible parties
– Example: “According to [Document], who owns the process for [specific action]?”

Classification Level
– Include relevant security or confidentiality classifications
– Example: “What are the requirements for [Classification Level] data?”

Stakeholder Scope
– Specify relevant stakeholders or approval authorities
– Example: “Which [stakeholder level] must approve [specific action]?”
</context_specification_requirements>

<question_quality_criteria>
– Questions must be specific enough that a vector database can match them to the relevant document chunk
– Questions should include key identifying terms, names, and context
– Questions should target concrete, actionable information
– Answers should provide complete context without referring back to document elements
</question_quality_criteria>

<output_format>
The question-answer-fact set should each be a short string in JSON format with the keys: “question”, “ground_truth_answer”, “fact”
</output_format>

<best_practices>
– Questions, answers, and facts should not refer to the subject entity as “it” or “they”, and instead refer to it directly by name
– Questions, answers, and facts should be individually unique to the document chunk, such that based on the question a new call to the retriever will address the correct document section when posing the ground truth question
– Facts should be represented in 3 or fewer words describing an entity in the <document>
– If there are units in the fact, the “fact” entry must provide multiple versions of the fact using <OR> as a delimiter. See <unit_variations> for examples.
<unit_variations>
– Dollar Unit Equivalencies: `1,234 million<OR>1.234 billion`
– Date Format Equivalencies: `2024-01-01<OR>January 1st 2024`
– Number Equivalencies: `1<OR>one`
</unit_variations>
</best_practices>

– Start your response immediately with the question-answer-fact set JSON, and separate each extracted JSON record with a newline.
</instructions>

<document>
{context_document}
</document>

Now, extract the question answer pairs and fact from the document excerpt according to your instructions, starting immediately with JSON and no preamble.”””

The generation output is provided as fact-wise JSONLines records in the following format, where elements in square brackets represent values from a line in Table 1.

{

“question”: “[Question]”,

“ground_truth_answer”: “[Ground Truth Answer]”,

“fact”: “[Fact]”

}

Here are a few examples of generated ground truth:

Question
Ground Truth Answer
Fact

What was Amazon’s total revenue growth in 2023?
Amazon’s total revenue grew 12% year-over-year from $514B to $575B in 2023.
12%<OR>$514B to $575B

How much did North America revenue increase in 2023?
North America revenue increased 12% year-over-year from $316B to $353B.
12%<OR>$316B to $353B

What was the growth in International revenue for Amazon in 2023?
International revenue grew 11% year-over-year from $118B to $131B.
11%<OR>$118B to $131B

How much did AWS revenue increase in 2023?
AWS revenue increased 13% year-over-year from $80B to $91B.
13%<OR>$80B to $91B

What was Amazon’s operating income improvement in 2023?
Operating income in 2023 improved 201% year-over-year from $12.2B to $36.9B.
201%<OR>$12.2B to $36.9B

What was Amazon’s operating margin in 2023?
Amazon’s operating margin in 2023 was 6.4%.
6.4%

Scaling ground truth generation with a pipeline
To automate ground truth generation, we provide a serverless batch pipeline architecture, shown in the following figure. At a high level, the AWS Step Functions pipeline accepts source data in Amazon Simple Storage Service (Amazon S3), and orchestrates AWS Lambda functions for ingestion, chunking, and prompting on Amazon Bedrock to generate the fact-wise JSONLines ground truth.

There are three user inputs to the step function:

A custom name for the ground truth dataset
The input Amazon S3 prefix for the source data
The percentage to sample for review.

Additional configurations are set by Lambda environment variables, such as the S3 source bucket and Amazon Bedrock Model ID to invoke on generation.

{

“dataset_name”: “YOUR_DATASET_NAME”,

“input_prefix”: “YOUR INPUT_PREFIX”,

“review_percentage”: “REVIEW PERCENTAGE”

}

After the initial payload is passed, a validation function assembles the global event payload structure in terms of system input and user input.

{

“system_input”:

{

“run_id”: “<AWS Step Function execution ID>”,

“input_bucket”: “<Input data Amazon S3 bucket>”,

“output_bucket”: “<Output data Amazon S3 bucket>”,

“output_document_chunks_prefix”: “<Amazon S3 bucket Prefix to store chunks>”,

“chunk_size”: “<Document chunk size>”,

“chunk_overlap”: “<Number of tokens that will overlap across consecutive chunks>”

},

“user_input”:

{

“dataset_name”: “<Dataset name>”,

“input_prefix”: “<Amazon S3 bucket prefix for ground truth generation data input data>”,

“review_percentage”: “<Percent of records to flag for human review>”

}

}

After validation, the first distributed map state iterates over the files in the input bucket to start the document ingestion and chunking processes with horizontal scaling. The resulting chunks are stored in an intermediate S3 bucket.
The second distributed map is the generation core of the pipeline. Each chunk generated by the previous map is fed as an input to the ground truth generation prompt on Amazon Bedrock. For each chunk, a JSONLines file containing the question-answer-fact triplets is validated and stored in an S3 bucket at the output prefix.
The following figure shows a view of the data structure and lineage from document paragraphs to the final ground truth chunk across the chunking and generation map states. The numbering between the two figures indicates the data structure present at each point in the pipeline. Finally, the JSONLines files are aggregated in an Amazon SageMaker Processing Job, including the assignment of a random sample for human review based on user input.

The last step of the pipeline is the aggregation step using a SageMaker Processing job. The aggregation step consists of concatenating the JSONLines records generated by every child execution of the generation map into a single ground truth output file. A randomly selected percentage of the records in the output file are sampled and flagged for review as part of a human-in-the-loop process.
Judging ground truth for FMEval question-answering evaluation
In this section, we discuss two key components of evaluating ground truth quality: human in the loop and applying an LLM as a Judge. Measuring ground truth quality is an essential component of the evaluation lifecycle.
Human-in-the-loop
The level of ground truth human review required is determined by the risk of having incorrect ground truth, and its negative implications. Ground truth review by use case SMEs can verify if critical business logic is appropriately represented by the ground truth. The process of ground truth review by humans is called human-in-the-loop (HITL), and an example the HITL process is shown in the following figure.
The steps of HTIL are:

Classify risk: performing a risk analysis will establish the severity and likelihood of negative events occurring as a result of incorrect ground truth used for evaluation of a generative AI use-case. Based on the outcome of the analysis, assign the ground truth dataset a risk level: Low, Medium, High or Critical. The table below outlines the relationship between event severity, likelihood, and risk level. See Learn how to assess the risk of AI systems for a deep dive on performing AI risk assessment.
Human review: Based on the assigned risk level, use-case expert reviewers examine a proportional amount of the use-case ground truth. Organizations can set acceptability thresholds for percentage of HITL intervention based on their tolerance for risk. Similarly, if a ground truth dataset is promoted from a low risk to a medium risk use case, an increased level of HITL intervention will be necessary.
Identify findings: Reviewers can identify any hallucinations relative to source data, challenges with information veracity according to their expertise, or other criteria set by the organization. In this post, we focus on hallucination detection and information veracity.
Action results: Reviewers can take business actions based on their judgement, such as updating and deleting records, or re-writing applicable source documents. Bringing in LLMOps SMEs to apply dataset curation best practices can also be an outcome.

Putting the risk table from Learn how to assess the risk of AI systems into action, the severity and likelihood of risks for a ground truth dataset validating a production chatbot with frequent customer use would be greater than an internal evaluation dataset used by developers to advance a prototype.

Likelihood

Severity
Rare
Unlikely
Possible
Likely
Frequent

Extreme
Low
Medium
High
Critical
Critical

Major
Very low
Low
Medium
High
Critical

Moderate
Very low
Low
Medium
Medium
High

Low
Very low
Very low
Low
Low
Medium

Very Low
Very low
Very low
Very low
Very low
Low

Next, we walk through the step-by-step process of conducting a human review for hallucination detection and information veracity. Human review is performed by comparing the ground truth chunk input to the LLM prompt to the generated question-answer-fact triplets. This view is shown in the following table.

Source data chunk
Ground truth triplets

Dear Shareholders: Last year at this time, I shared my enthusiasm and optimism for Amazon’s future. Today, I have even more. The reasons are many, but start with the progress we’ve made in our financial results and customer experiences, and extend to our continued innovation and the remarkable opportunities in front of us. In 2023, Amazon’s total revenue grew 12% year-over-year (“YoY”) from $514B to $575B. By segment, North America revenue increased 12% Y oY from $316B to $353B, International revenue grew 11% YoY from $118B to $131B, and AWS revenue increased 13% YoY from $80B to $91B.
{“question”: “What was Amazon’s total revenue growth in 2023?”, “ground_truth_answer”: “Amazon’s total revenue grew 12% year-over-year from $514B to $575B in 2023.”, “fact”: “12%<OR>$514B to $575B”} {“question”: “How much did North America revenue increase in 2023?”, “ground_truth_answer”: “North America revenue increased 12% year-over-year from $316B to $353B.”, “fact”: “12%<OR>$316B to $353B”} {“question”: “What was the growth in International revenue for Amazon in 2023?”, “ground_truth_answer”: “International revenue grew 11% year-over-year from $118B to $131B.”, “fact”: “11%<OR>$118B to $131B”}

Human reviewers then identify and take action based on findings to correct the system. LLM hallucination is the phenomenon where LLMs generate plausible-sounding but factually incorrect or nonsensical information, presented confidently as factual. Organizations can introduce additional qualities for evaluating and scoring ground truth, as suited to the risk level and use case requirements.
In hallucination detection, reviewers seek to identify text that has been incorrectly generated by the LLM. An example of hallucination and remediation is shown in the following table. A reviewer would notice in the source data that Amazon’s total revenue grew 12% year over year, yet the ground truth answer hallucinated a 15% figure. In remediation, the reviewer can change this back to 12%.

Source data chunk
Example hallucination
Example hallucination remediation

In 2023, Amazon’s total revenue grew 12% year-over-year (“YoY”) from $514B to $575B.
{“question”: “What was Amazon’s total revenue growth in 2023?”, “ground_truth_answer”: “Amazon’s total revenue grew 15% year-over-year from $514B to $575B in 2023.”, “fact”: “12%<OR>$514B to $575B”}
{“question”: “What was Amazon’s total revenue growth in 2023?”, “ground_truth_answer”: “Amazon’s total revenue grew 12% year-over-year from $514B to $575B in 2023.”, “fact”: “12%<OR>$514B to $575B”}

In SME review for veracity, reviewers seek to validate if the ground truth is in fact truthful. For example, the source data used for the ground truth generation prompt might be out of date or incorrect. The following table shows the perspective of an HITL review by a domain SME.

Source data chunk
Example SME review
Example hallucination remediations

Effective June 1st, 2023, AnyCompany is pleased to announce the implementation of “Casual Friday” as part of our updated dress code policy. On Fridays, employees are permitted to wear business casual attire, including neat jeans, polo shirts, and comfortable closed-toe shoes.
“As an HR Specialist, this looks incorrect to me. We did not implement the Casual Friday policy after all at AnyCompany – the source data for this ground truth must be out of date.”

Delete Incorrect Ground Truth
Update Source Data Document
Other use case specific actions

Traditional machine learning applications can also inform the HITL process design. For examples of HITL for traditional machine learning, see Human-in-the-loop review of model explanations with Amazon SageMaker Clarify and Amazon A2I. 
LLM-as-a-judge
When scaling HITL, LLM reviewers can perform hallucination detection and remediation. This idea is known as self-reflective RAG, and can be used to decrease—but not eliminate—the level of human effort in the process for hallucination detection. As a means of scaling LLM-as-a-judge review, Amazon Bedrock now offers the ability to use LLM reviewers and to perform automated reasoning checks with Amazon Bedrock Guardrails for mathematically sound self-validation against predefined policies. For more information about implementation, see New RAG evaluation and LLM-as-a-judge capabilities in Amazon Bedrock and Prevent factual errors from LLM hallucinations with mathematically sound Automated Reasoning checks (preview).
The following figure shows an example high-level diagram of a self-reflective RAG pattern. A generative AI application based on RAG yields responses fed to a judge application. The judge application reflects on whether responses are incomplete, hallucinated, or irrelevant. Based on the judgement, data is routed along the corresponding remediation.

The golden rule in implementing HITL or LLM-as-a-judge as part of ground truth generation is to make sure the organization’s review process aligns with the accepted risk level for the ground truth dataset.
Conclusion
In this post, we provided guidance on generating and reviewing ground truth for evaluating question-answering applications using FMEval. We explored best practices for applying LLMs to scale ground truth generation while maintaining quality and accuracy. The serverless batch pipeline architecture we presented offers a scalable solution for automating this process across large enterprise knowledge bases. We provide a ground truth generation prompt that you can use to get started with evaluating knowledge assistants using the FMEval Factual Knowledge and QA Accuracy evaluation metrics.
By following these guidelines, organizations can follow responsible AI best practices for creating high-quality ground truth datasets for deterministic evaluation of question-answering assistants. Use case-specific evaluations supported by well-curated ground truth play a crucial role in developing and deploying AI solutions that meet the highest standards of quality and responsibility.
Whether you’re developing an internal tool, a customer-facing virtual assistant, or exploring the potential of generative AI for your organization, we encourage you to adopt these best practices. Start implementing a robust ground truth generation and review processes for your generative AI question-answering evaluations today with FMEval.

About the authors
Samantha Stuart is a Data Scientist with AWS Professional Services, and has delivered for customers across generative AI, MLOps, and ETL engagements. Samantha has a research master’s degree in engineering from the University of Toronto, where she authored several publications on data-centric AI for drug delivery system design. Outside of work, she is most likely spotted playing music, spending time with friends and family, at the yoga studio, or exploring Toronto.
Philippe Duplessis-Guindon is a cloud consultant at AWS, where he has worked on a wide range of generative AI projects. He has touched on most aspects of these projects, from infrastructure and DevOps to software development and AI/ML. After earning his bachelor’s degree in software engineering and a master’s in computer vision and machine learning from Polytechnique Montreal, Philippe joined AWS to put his expertise to work for customers. When he’s not at work, you’re likely to find Philippe outdoors—either rock climbing or going for a run.
Rahul Jani is a Data Architect with AWS Professional Service. He collaborates closely with enterprise customers building modern data platforms, generative AI applications, and MLOps. He is specialized in the design and implementation of big data and analytical applications on the AWS platform. Beyond work, he values quality time with family and embraces opportunities for travel.
Ivan Cui is a Data Science Lead with AWS Professional Services, where he helps customers build and deploy solutions using ML and generative AI on AWS. He has worked with customers across diverse industries, including software, finance, pharmaceutical, healthcare, IoT, and entertainment and media. In his free time, he enjoys reading, spending time with his family, and traveling.

Hugging Face’s SmolAgents framework provides a lightweight and efficient way to build AI agents that leverage tools like web search and code execution. In this tutorial, we demonstrate how to build an AI-powered research assistant that can autonomously search the web and summarize articles using SmolAgents. This implementation runs seamlessly, requiring minimal setup, and showcases the power of AI agents in automating real-world tasks such as research, summarization, and information retrieval.

Copy CodeCopiedUse a different Browser!pip install smolagents beautifulsoup4

First, we install smolagents beautifulsoup4, which enables AI agents to use tools like web search and code execution, and BeautifulSoup4, a Python library for parsing HTML and extracting text from web pages.

Copy CodeCopiedUse a different Browserimport os
from getpass import getpass

# Securely input and store the Hugging Face API token
os.environ[“HUGGINGFACEHUB_API_TOKEN”] = getpass(“Enter your Hugging Face API token: “)

Now, we securely input and store the Hugging Face API token as an environment variable. It uses getpass() to prompt users to enter their token without displaying it for security. The token is then stored in os.environ[“HUGGINGFACEHUB_API_TOKEN”], allowing authenticated access to Hugging Face’s Inference API for running AI models.

Copy CodeCopiedUse a different Browserfrom smolagents import CodeAgent, DuckDuckGoSearchTool, HfApiModel

# Initialize the model WITHOUT passing hf_token directly
model = HfApiModel()

# Define tools (DuckDuckGo for web search)
tools = [DuckDuckGoSearchTool()]

# Create the agent
agent = CodeAgent(tools=tools, model=model, additional_authorized_imports=[“requests”, “bs4”])

Now, we initialize an AI-powered agent using the SmolAgents framework. It sets up HfApiModel() to load a Hugging Face API-based language model, automatically detecting the stored API token for authentication. The agent is equipped with DuckDuckGoSearchTool() to perform web searches. Also, CodeAgent() is instantiated with tool access and authorized imports, such as requests for making web requests and bs4 for parsing HTML content.

Copy CodeCopiedUse a different Browser# Example query to the agent:
query = “Summarize the main points of the Wikipedia article on Hugging Face (the company).”

# Run the agent with the query
result = agent.run(query)

print(“nAgent’s final answer:n”, result)

Finally, we send a query to the AI agent, asking it to summarize the main points of the Wikipedia article on Hugging Face. The agent.run(query) command triggers the agent to perform a web search, retrieve relevant content, and generate a summary using the language model. Finally, the print() function displays the agent’s final answer, concisely summarizing the requested topic.

Sample Output

Following this tutorial, we have successfully built an AI-powered research assistant using Hugging Face SmolAgents that can autonomously search the web and summarize articles. This implementation highlights the power of AI agents in automating research tasks, making it easier to retrieve and process large amounts of information efficiently. Beyond web search and summarization, SmolAgents can be extended to various real-world applications, including automated coding assistants, personal task managers, and AI-driven chatbots.

Here is the Colab Notebook for the above project. Also, don’t forget to follow us on Twitter and join our Telegram Channel and LinkedIn Group. Don’t Forget to join our 80k+ ML SubReddit.

Recommended Read- LG AI Research Releases NEXUS: An Advanced System Integrating Agent AI System and Data Compliance Standards to Address Legal Concerns in AI Datasets
The post Step by Step Guide to Build an AI Research Assistant with Hugging Face SmolAgents: Automating Web Search and Article Summarization Using LLM-Powered Autonomous Agents appeared first on MarkTechPost.