February 2025 Summaries

An AI platform is a unified technology stack for prototyping, developing, releasing, and maintaining GenAI-enabled applications, comprising several components such as general-purpose or bespoke LLMs, retrieval-augmented generation techniques like RAG, data ingestion, user-facing agents, request/response workflows, support for multiple LLMs, and abstraction on top of an LLM. While not necessary, using an AI platform can reduce the work of eliminating universal features, increase production readiness, increase the pool of GenAI app developers, provide advanced features for extracting data from structured, unstructured, and non-structured sources, offer integration with different LLMs, support for switching out LLMs mid-project, and enable composability and orchestration. A good AI platform should make developers more productive, reduce ramp-up time, support creating no-code and low-code solutions, provide a visual builder, and integrate with multiple systems.

In this post, we'll showcase advanced Storage-attached indexing (SAI) capabilities in Apache Cassandra 5.0 and demonstrate how to convert Solr schema fields to corresponding SAI index options, particularly with index analyzers, tokenizers, and filters. We're using a notional application that stores information about movies and uses Solr for search. The movie data is partitioned by the movie ID, and we create a regular SAI index on the title column. However, due to the StandardTokenizer and LowerCaseFilter in the Solr schema, this results in non-exact matches when searching for the term "Extraction". To overcome this, we use the built-in STANDARD analyzer in Cassandra to split the text into words and convert them into lowercase. This allows us to get hits for movies with more than just the search term "Extraction" in the title. We can also use other generic analyzers like simple, whitespace, stop, and lowercase, as well as specific analyzers for over 30 languages. Additionally, we can add a stemming filter like PorterStemFilter to normalize terms into their base words. SAI supports the use of CONTAINS and CONTAINS KEY operators when querying collections, allowing us to search for action movies by genre or Chris Hemsworth in movies.

The Cassandra Data Migrator (CDM) is a tool used to move data between Apache Cassandra and DataStax Astra DB, it works with the Apache Spark framework to transfer large amounts of data without intermediate storage. The CDM can be run as a Docker container or built from its JAR file, it requires configuration through the `cdm.properties` file specifying connection details, keyspace, table names, and migration settings. In contrast, the DataStax Bulk Loader (DSBulk) is a versatile tool that can work with any database operating on the CQL protocol standard, including Cassandra, and provides functionality for exporting, importing, and counting rows in Cassandra tables. DSBulk can be configured through command line flags and has been used successfully in many Cassandra migrations, but may introduce load onto the origin Cassandra cluster during export operations. Both tools have their strengths and weaknesses, and it is recommended to assess the cluster's resources before migration to minimize impact on production usage.

IBM plans to acquire DataStax to accelerate the use of artificial intelligence (AI) and NoSQL data at scale, addressing a common challenge faced by enterprises adopting AI: leveraging their enterprise data estate to power AI agents and large language models. Industry leaders such as FedEx, Capital One, and Verizon have chosen DataStax and Apache Cassandra to power critical applications. The acquisition will bring together scalable, secure, and accurate production AI capabilities with IBM's existing technologies, including the hybrid vector database and open-source AI tool Langflow. DataStax's NoSQL products, such as Astra DB, will continue to be available, serving customers who rely on them for scale-out data delivery. The acquisition is subject to close of the transaction and regulatory approval, and DataStax remains committed to its mission of supporting customers with inspired leadership and open-source principles.

The Apache Cassandra project, an open-source database management system, is being acquired by IBM's DataStax subsidiary, marking a significant moment in the community's journey. The acquisition signals a strategic bet on Cassandra's future, particularly for hybrid-cloud capabilities and advanced AI integration. DataStax, which has been supporting the project since its inception, will continue to contribute to Cassandra and ensure a seamless transition for customers. With IBM's resources combined with Cassandra's momentum, new possibilities are opening up, including advancements in usability enhancements such as AI features, ACID transactions, and more. The acquisition aligns with DataStax's commitment to open-source principles and its dedication to serving customers who rely on Cassandra for mission-critical applications. As the project moves forward, it aims to turn into a database that serves a wide range of needs without sacrificing its core strengths, including linear scalability, global replication, and reliability.

DSE Search is a powerful search engine built on Apache Solr, but its limitations, such as the two billion document limit per shard index and write-amplification, can hinder performance. DataStax's Storage-attached indexing (SAI) provides a more flexible and performant alternative for equality filters on indexed columns, without requiring full-text search capabilities. By using SAI indexes with Astra DB, developers can replace complicated Solr queries with faster lookups, such as filtering movies by title or release year, with syntax familiar to developers. SAI offers efficient retrieval of results due to its unique indexing mechanism and supports range queries, enabling developers to build high relevancy and low latency applications.

Graph-based retrieval-augmented generation (RAG) can yield better results than traditional vector search in certain scenarios, especially when dealing with highly interlinked sources such as technical documents or Web pages. LangChain and a vector database like Astra DB can simplify adding this technique to your GenAI data stack, allowing developers to get started with just a few lines of code. RAG takes information from a GenAI query and supplements it with domain-specific context that's both relevant and current, using a graph structure and graph traversal to compensate for potential limitations of vector search. Knowledge graphs can be built automatically without requiring extensive manual definition and tweaking, making them a versatile tool for searching a wide variety of content. With LangChain and Astra DB, developers can generate a content-centric graph that stores information in a hybrid graph/vector format, allowing for both similarity (vector search) and graph traversal search methods. This approach enables the creation of highly performant GenAI apps with high relevancy and low latency.

With the OpenAI Realtime API, developers can build speech-to-speech applications that interact directly with a generative AI model by speaking to it. The API enables the creation of tools that the model can use to execute functions and extend its capabilities. A developer built an example using Twilio's platform, connecting a phone call to GPT-4o with Node.js or Python, but wanted to explore more possibilities. To achieve this, they extended the original assistant into an agent that can choose to use tools to augment its response. The agent uses retrieval-augmented generation (RAG) with Astra DB to gather up-to-date data and store it in a database. This allows the model to search the database for additional information, providing more accurate responses. The developer created a tool that enables the model to perform vector searches against the collection, returning relevant text chunks as context. The agent now provides a new way to connect with the Taylor Swift bot, enabling users to chat and ask questions about the singer-songwriter using voice or phone calls. The combination of Twilio, OpenAI, and Astra DB creates a powerful agent that can leverage tools like RAG to provide more accurate responses.

The adoption of AI agents is being hindered by concerns about trust, safety, and governance. Nearly half of respondents to a recent survey expressed concerns about the ethical implications of deploying AI agents in their industries, with 48.3% citing trust and safety as top barriers to adoption. However, many respondents also saw benefits from using AI agents, including cost savings, speed improvements through augmentation, and acceleration of human productivity. Despite these benefits, 45.5% of respondents remained unsure if agent frameworks adequately address the real-world challenges of building agentic applications for production. To help mitigate these risks, organizations are looking to implement guardrails, humans remaining in the loop, traceability, and evaluations. A low-code visual development environment called Langflow is being used to simplify the creation of multi-agent applications, and a hackathon is planned to bring together developers to build with the latest AI tooling.

Ollama is an open-source platform that enables developers to run AI models securely on their own machine without sharing data with external providers, providing a more secure and cheaper alternative to public model providers. It allows users to download and install the necessary software, navigate to the models section, select tools that support tool calling, and use visual IDEs like Langflow to build generative and agentic AI flows in a low-code or no-code environment. With Ollama, developers can confidently run models on their hardware, knowing that their data is safe and secure, without exposing it to public model providers. The platform supports local tool calling and agent interactions, making it suitable for running smaller machines. Users can experiment with their inputs, adjust the model's temperature settings, and integrate their Ollama agents into applications via API to enable full agentic capability.

DataStax is hosting a Valentine's Day giveaway, where participants can generate a personalized poem about themselves based on their GitHub username through a generative AI application. The most-liked post will win a pair of DataStax purple BOSE wireless earbuds. The giveaway is open to individuals aged 18 and older, excluding employees and immediate family members, and runs from February 10 to February 13, 2025. Participants must share their poem on X and tag DataStaxDevs for it to be eligible. Winners will be notified within 48 hours of selection and must respond within 72 hours or forfeit the prize. The giveaway is subject to terms and conditions, which include restrictions on eligibility, entry requirements, and winner notification.

You can create an AI agent in Langflow using Composio as a tool, an integration platform that gives developers access to hundreds of tools like GitHub, Salesforce, and Google. With Composio, you don't have to manage APIs for performing actions like creating, deleting, or updating a Google Calendar event; you just need to go through Composio and the work is done for you. To get started, create a new account with Composio and obtain an API key, then set up a Google calendar integration by following the steps in the platform's dashboard. You can then use this API key in Langflow to build a simple AI agent that interacts with your Google Calendar using the `GOOGLECALENDAR` app name. The agent uses the gpt-4o-mini model from OpenAI, but you can choose to use other models if preferred. With Composio and Langflow, you can easily connect multiple tools and perform actions without having to configure APIs yourself, and explore many integrations using Langflow's component-based visual app-building interface.

Knowledge graphs are data structures that represent entities, their relationships, and organizing principles, providing added value over relational or vector data by extracting multiple facts from a single source, representing structured, unstructured, and semi-structured data, and correlating relationships across different sources. They can also retrieve knowledge several steps away from the original entities in question in a single query, making them useful for applications like GenAI, search engines, real-time fraud detection, and product recommendation engines. Knowledge graphs are implemented as either triple stores or property graphs, with property graphs being easier to use and faster to query compared to RDF. They can be used to implement retrieval-augmented generation (RAG) in GenAI apps by establishing connections between sources and providing more relevant context for LLM queries.

Generative AI applications will remain a primary focus for enterprises in the next year or two, but they may not fully realize their potential until agentic AI is integrated into enterprise architectures. Agentic AI employs autonomous agents to adapt to dynamic environments and create results independently. When applied to business process workflows, it can replace static processes with dynamic automation systems. Agentic AI enables tools use, decision-making, planning, and reasoning functions in agents, which can work together to achieve goals. Frameworks are necessary to orchestrate these agents, enabling them to provide additional information or refine each other's outputs. This approach allows developers to build agentic processes by mixing and matching components, such as models, user input, and business services. Tools like Langflow simplify the process of building agentic systems through visual, low-code builders. Agentic workflows bring together enterprise data, AI, and APIs, forming systems of automation that empower domain experts to scale their abilities and improve enterprises through AI.