April 2025 Summaries

Astra DB now supports hybrid search, which combines vector search and BM25 keyword search to improve the accuracy of search results by up to 45%. Hybrid search uses a reranking model to combine the results from both searches and returns the most relevant results. Astra DB can perform hybrid search using its built-in NVIDIA NeMo Retriever reranking microservices, which allows developers to create vector embeddings for their unstructured content and store them in a database. The hybrid search process involves creating a collection with settings that define how to create vectors and handle keyword searches. Developers can ingest data into the collection and perform vector and hybrid searches against it. Hybrid search improves relevance by combining the strengths of both vector and keyword search, making it useful for retrieval-augmented generation (RAG) applications. By using hybrid search, developers can improve the accuracy of their RAG application's output, reducing the likelihood of inaccurate or hallucinated responses.

DataStax has joined the OpenSearch Software Foundation to drive open innovation in data search and analytics, aligning with its history of supporting open source database software. The company has been working with OpenSearch for some time and recently announced an integration with JVector, a vector search engine that provides flexible information retrieval capabilities. DataStax considers OpenSearch the future of enterprise search, particularly as it expands its self-managed search offerings, and is excited to support users in getting the most out of their enterprise data estates. Kiyu Gabriel, General Manager for Search & AI at DataStax, drives innovations in artificial intelligence solutions and real-time AI analytics, bringing his expertise in enterprise architecture, development, and engineering to the partnership.

Astra DB supports vector-only, hybrid, and graph RAG searches on a single vector collection, enabling users to switch between search strategies as needed. The database stores embeddings, content, and metadata together, facilitating flexible data management. Astra DB offers various development tools, including the Data API, Langflow, and Python libraries like LangChain Graph Retriever, making it easy for developers to start building AI applications. Charna Parkey, a seasoned tech executive and expert in B2B AI products, emphasizes the importance of using AI in transparent and accountable ways to combat systemic oppression. She hosts the podcast Open Source Data, exploring the intersection of AI development, regulation, academia, and industry, with the goal of mitigating risk and driving forward responsible AI system design.

The biggest problem with most generative AI apps isn't the model, but rather the data layer, which can lead to vague or misleading results if the retrieval pipeline can't deliver accurate context. Most search architectures weren't designed for GenAI and rely on vector search alone, which is not enough to achieve accuracy. To build GenAI apps that perform in production, a hybrid approach is necessary, combining vector search with reranking techniques that evaluate retrieved content against user queries. This approach improves accuracy by up to 45 percent compared to vector-only search and provides a more accurate context for domains like healthcare, legal, and customer support.

Langflow is a software development tool that allows developers to create webhooks, which expose an endpoint to the developer's workflow. To set up a webhook in Langflow, developers need to bring in the Webhook component and obtain a cURL command, test the setup by sending a POST request to the webhook endpoint, and confirm that the data was sent successfully. The Parser component can then be used to extract specific fields from the incoming data using a template field. Developers can use this data in their workflow by connecting the output of the Webhook component to the Data or DataFrame input of the Parser component. To dynamically interact with Langflow from an application or external service, developers need to define the webhook URL, integrate the webhook call, and implement error handling. They can also use services like Ngrok or Render to deploy their Langflow instance to access the webhook URL externally. By harnessing the power of webhooks in Langflow, developers can create more dynamic and interactive workflows that can be triggered by external events or services.

DataStax has partnered with NVIDIA to integrate their NeMo microservices into its data and AI platforms, including Astra DB, DataStax Hyper-Converged Database, and the DataStax AI Platform. This partnership aims to provide scalable RAG (retrieval-augmented generation) and hybrid search functionality for GenAI applications. The integration of NeMo microservices has resulted in significant improvements in performance, cost-effectiveness, and latency, with sub-12ms latency on NVIDIA H100 GPUs. DataStax has also launched Astra Vectorize, which enables embedding generation at the database tier, and DataStax Hyper Converged Data (HCD) Platform, which brings OpenSearch and Apache Pulsar to enable GenAI and vector search capabilities. The partnership with NVIDIA has enabled DataStax to create a platform that harnesses the full potential of its data through AI, turning information into intelligence and intelligence into action.

The emergence of retrieval-augmented generation (RAG) has revolutionized the accuracy and contextual relevance of generative AI outputs, particularly in applications where traditional approaches have limitations such as missing nuanced contextual relationships or structured associations between documents. Recent advancements like Unstructured and the Graph Retriever library have simplified graph-based RAG by providing push-button transformation of unstructured documents into structured, graph-ready data, eliminating the need for dedicated graph databases. This structured approach provides superior context navigation, enabling applications to fetch documents related not merely semantically but based explicitly on relationships and entities present in the metadata. Unstructured's extensive and extensible metadata out of the box enhances the accuracy of retrieval, while its declarative approach enables non-developers to build workflows seamlessly, accelerating development and ensuring efficient execution at scale. The Graph Retriever library builds on LangChain vector stores, enabling dynamic graph construction from structured metadata, enhancing retrieval flexibility, context-awareness, and precision without additional complex infrastructure.

Achieving accuracy in AI applications is crucial for driving critical business processes and meeting ambitious goals, as demonstrated by the world's largest companies and emerging startups that have quickly gotten up to speed with Langflow and Arize. The key to achieving accuracy lies in understanding four pillars of AI accuracy: data preparation, retrieval, monitoring and evaluation, and reranking and optimization. Next week, experts will share expert tips, code, and practical guidance on how to build accurate AI apps by avoiding hallucinations, reducing drift, and building AI that gets it right every time. The knowledge shared during Accuracy Week will go a long way in improving production generative AI applications, with highlights including a livestream featuring Rich Young, a seasoned product and technology leader, and a chance to learn from Ed Anuff's 30 years of experience as a product and technology leader at top companies such as Google and Apigee.

The DataStax team has been reflecting on its partnership with Google, highlighting recent collaborations that have enabled enterprises and developers to take advantage of the latest AI innovations. The integration with Gemini Code Assist tools allows developers to seamlessly access their database structures, create data models, and get query assistance within their IDE. This collaboration streamlines development workflows and enhances productivity for developers who rely on Astra DB and Langflow. Customers have reported significant improvements in performance and scalability using Astra DB, including a 90% reduction in response times for WealthAPI's next-generation wealth management app. The DataStax team is contributing to the definition of Google's open Agent2Agent (A2A) interoperability protocol, which enables intelligent agents to communicate and collaborate securely without requiring access to each other's internal architectures. Additionally, DataStax has joined Google Cloud's Startup Perks program as a launch partner, offering exclusive discounts and top-tier technology solutions to help startups build and scale efficiently.

Astra DB Genkit plugin is a new integration that makes it easy to build retrieval-augmented generation (RAG) applications in Genkit using Astra DB for accurate, low-latency vector search. This plugin integrates Astra DB into the Genkit framework as an indexer and retriever, enabling developers to efficiently embed, index, and retrieve data within their Genkit applications using Astra DB. The plugin is open source and available on GitHub, providing a unified API for working with LLMs, easy monitoring of applications in production, and making it easy to build RAG applications. With the Astra DB Genkit plugin, developers can quickly integrate a scalable vector database into their Genkit AI applications, reducing the challenges associated with building AI-powered applications from data infrastructure management to the orchestration of intricate AI features.

The Firebase Genkit framework is a powerful tool for building generative AI applications, providing essential tools such as models, prompts, indexers, retrievers, flows, traces, and evaluations. To build a Retrieval-Augmented Generation (RAG) application with Genkit and Astra DB, developers need to create a database and collection within Astra DB, install Genkit and related dependencies into their Node.js application, configure Genkit with their Astra DB credentials, and then start creating flows. These flows can be used to ingest data from a URL, retrieve context from the database, generate responses based on the context, and perform vector search using the Astra DB retriever. The process of building an RAG application takes fewer than 100 lines of code, making it relatively straightforward, thanks to Firebase Genkit's ease of use and powerful features such as tracing that helps track down bugs in flows.

Agentic workflows are networks of autonomous GenAI agents that perform complex tasks with minimal human intervention, making decisions and taking actions on a user's behalf. Building an agentic workflow is essentially taking a componentized, modular approach to GenAI agents, bringing benefits such as greater accuracy, better performance, and improved reusability. Agentic workflows can be built easily using tools like Langflow, which enables visually composing, testing, and deploying complex GenAI apps from pre-built components, cutting development time down from weeks to days. These workflows can harness the power of large language Models (LLMs) themselves to assist in breaking down work, parceling out subtasks to agents that specialize in answering certain queries. Agentic AI is seen as a self-driving car for AI agents - an intelligent and flexible AI that can adapt its approach to different user queries and changing circumstances.

When building a retrieval-augmented generation (RAG) app, you need to prepare your data by creating vector embeddings in various ways such as locally, via API, via a framework or with Astra DB's Vectorize. Pre-trained embedding models like Sentence Transformers and all-MiniLM-L6-v2 can be used to generate vector embeddings. Local embedding models are useful for experimentation on laptops or hardware acceleration, while APIs provided by services like OpenAI, Google and Cohere offer an alternative option. Frameworks like LangChain and LlamaIndex provide standardized interfaces that abstract the complexities of embedding models and APIs. Astra Vectorize enables Astra DB to automatically generate vector embeddings as documents are inserted or queries are performed, simplifying code maintenance, improving performance and efficiency.

### Storage-attached indexing is being introduced in Apache Cassandra 5.0, allowing for more flexible and performant query patterns with less coding required. DataStax has been working on this feature, SAI (Storage-Attached Indexing), for several years and it's now deployed in their Cassandra-as-a-service Astra DB, showing high reliability and performance. SAI is better than other indexing methods available for CQL, providing more functionality at a fraction of the storage footprint, using less disk space than Solr implementation, and having significantly better throughput and lower latency compared to 2i and Solr. SAI indexes can be created on any column in a table except partition key columns, allowing queries on any node in any data center without the need for separate DCs or resources. SAI is built on Apache Lucene so it works with built-in analyzers to extract index terms from text like Solr, including standard and simple tokenizers. SAI enables semantic search with natural language processing (NLP) and machine learning algorithms to provide more accurate results. It also supports vector search which enables generative AI capabilities by extracting the essence of unstructured data using LLMs and generating vector embeddings. This allows for quick location of top N rows in a table that are most similar to a user query, all this can be done with just five lines of code.

This is an announcement from Apache Cassandra about its new comprehensive data modeling series for 2025, designed to provide developers and architects with the latest knowledge on building modern applications. The series will cover Cassandra 5.0 as the basis for building an application, starting with foundational principles that remain unchanged even in newer versions of Cassandra. It will guide learners through a progressive learning journey using KillrVideo, a YouTube-like application, exploring revolutionary storage-attached indexes, enhanced data types and functions, data protection and governance, and practical migration strategies to modernize existing implementations. The series offers concrete techniques for implementing immediate improvements to data models, resulting in more intuitive, maintainable, and compliant applications that are optimized for performance at any scale.

The development of large language models has led to a shift in the way developers approach AI applications, with a focus on creating agents that can reason, adapt, and take action. However, this shift has also created challenges, as models become increasingly probabilistic black boxes, making it difficult to debug and inspect their behavior. Langflow is an agentic model system that addresses these challenges by providing a modular approach to agent design, allowing developers to build atomic tools and functions, combine them into purpose-built agents, and use orchestration to route tasks to the right agent at the right time. This approach enables scalability while maintaining control and visibility, making it easier to debug and evaluate the behavior of the model. Langflow also provides a visual interface for building agents as tools, creating nested agent chains, and routing tasks based on context, mirroring both software engineering best practices and human collaboration patterns.