January 2025 Summaries

Master DeepSeek & Langflow Retrieval-augmented generation (RAG) is a reliable method for generating additional context for generative AI apps, but its effectiveness can vary depending on the type of data used. When source information consists of heavily interlinked documents, a knowledge graph can provide more accurate results than a vector database. A knowledge graph represents information as nodes and relationships between them, allowing for semantically relevant results to be found by traversing links between nodes. To build and store a knowledge graph, there are two standard models: Resource Description Framework (RDF) and property graphs. RDF-formatted graphs can be stored in an RDF triplestore, which offers benefits such as handling complexity at scale and being less costly to implement. In contrast, property graphs are typically stored in a graph database, offering flexibility but requiring additional development effort. However, leveraging existing vector databases like Astra DB or using tools like LangChain's GraphVectorStore can simplify the process without adding significant lift to the GenAI app infrastructure. By utilizing these approaches, developers can support GraphRAG without significantly changing their existing stack, resulting in more accurate results and improved performance.

DataStax's Langflow is a technical preview of its flagship product, aiming to democratize generative AI by providing a fully open path to creating GenAI flows and agents for rapid prototyping and proofs of concept. With Langflow.new, developers can immediately get started with Langflow, discovering its value in building RAG pipelines, agentic flows, and more. The platform features components called tools that work similarly to functions in programming, where input parameters are generated and supplied by a language model, and their outputs are returned to the model. These tools are used to execute user prompts, such as calculating sums or converting currencies, and the platform's glue (runtime) Langflow enables seamless interaction between the language model, tools, and user inputs. With Langflow.new, users can experiment with agentic flows and tools, download their workflows, and use them with deployed Langflow instances, either hosted by DataStax or self-hosted. The goal is to reduce friction and increase access to meaningful GenAI flows, allowing users to create high-relevance and low-latency RAG apps using AI.

The Langflow client is a JavaScript library that allows developers to integrate the power of Langflow into their Node.js applications, enabling them to build generative AI flows and agents with ease. The client provides an easy-to-use API for interacting with the Langflow API, which can be used to make calls directly or through the client. The client supports both DataStax-hosted and self-hosted versions of Langflow, and requires an OpenAI API key to run. Developers can use the client to initialize a client, run flows, and access the flow response, with features such as tweakable components and flexible settings. The library is open-source and available on GitHub, and provides a complete data API and out-of-the-box integrations for building production RAG apps with high relevancy and low latency.

2025 is the year of AI agents, according to a Google paper from October 2024, an AI agent is software that consists of three tiers. Langflow is a low-code development environment that treats agents as first-class citizens, allowing users to create and run simple agents with ease. By using the "Simple Agent" template in Langflow, users can connect components such as a calculator and web browser to solve real-world problems like currency conversion. The agent's knowledge and capabilities can be extended through various means, including API integrations, and it can be consumed by other applications via network requests. Astra DB provides a complete data API for building production RAG apps with high relevancy and low latency.

WealthAPI has developed an AI-driven solution to identify recurring payments in bank transaction histories, addressing the challenges of traditional methods which struggled with scaling and relied on exact matches, missing subtle differences. The system uses vector embeddings to group transactions into recurring payment patterns, ensuring accuracy even when entries contain wording differences. WealthAPI's architecture is powered by DataStax Astra DB, a database platform purpose-built for modern, scalable, and AI-integrated workflows. Astra DB ensures scalability and responsiveness at each stage of the process, adhering to strict data security measures to protect end users' transactions. The system can efficiently search for similar transactions even with minor variations in details, thanks to its ability to store embeddings and run simple similarity searches. WealthAPI's use of Astra DB demonstrates how advanced database technology can drive innovation in financial analytics, delivering faster, smarter solutions to its clients.

Leveraging large language models (LLMs) in critical business processes, customer-facing agents, or compliance-driven scenarios requires accurate, contextual, and verifiable information to ensure accuracy. Establishing a reliable ground-truth dataset, which includes questions and validated answers representing the correct responses for a given domain, is key. However, generating such a dataset can be costly, complex, and labor-intensive. A new toolkit enables organizations to automate this process by harnessing the power of LLMs themselves and using an image-based workflow that preserves the original layout and structure of documents. This approach delivers more accurate and reliable ground truth datasets, faster than traditional methods, by considering every element, including table cells, images, captions, and layout nuances. By anchoring LLMs in authoritative sources, organizations can ensure answers are both domain-relevant and contextually precise, reducing hallucinations, fostering trust, and supporting compliance. The toolkit also provides a streamlined workflow for creating and refining ground truth datasets, making it easier to build production-grade language models.

Graph retrieval-augmented generation (RAG) is a powerful tool that integrates multiple data sources to produce more accurate results by leveraging a knowledge graph to reveal hidden relationships and structures in documents. This new perspective aids in document visualization and enhances RAG performance for large language model queries, particularly in capturing hierarchical or relational connections within documents. Unstructured excels at parsing and interpreting different types of documents, extracting both content and metadata, which lays the groundwork for building a graph that captures document hierarchies and connections often missed by purely semantic embeddings. Astra DB complements this by providing a scalable, hybrid graph and vector store, ensuring scalability as data grows. By combining these tools, users can create a rich, interconnected perspective of their documents, enabling smarter, more context-aware results for LLMs. Graph RAG provides a new way to visualize and organize documents, uncovering hidden relationships and structures, which enables more effective understanding of the data while enhancing traditional RAG performance.

Scraping web pages is a useful way to fetch content for retrieval-augmented generation (RAG) applications, but parsing the content from a web page can be challenging due to irrelevant information like headers and footers. Mozilla's open-source library Readability.js is a helpful tool for extracting just the important parts of a web page, allowing developers to remove irrelevant content and return high-quality results. By using Readability.js in a data pipeline, developers can strip out unnecessary content and focus on the main subject of the page, making it easier to build RAG-powered applications with high relevancy and low latency. The library is battle-tested, powering Firefox's reader mode, and can be used directly or integrated into frameworks like LangChain.js for more complex applications.

The latest commercial vector embedding models have been released, including proprietary and open-source options from prominent vendors such as Gemini, OpenAI, Jina, Cohere, Voyage, Stella, ModernBert Embed, and TabFQuAD/Shift Project. The models were tested on the ViDoRe image search benchmark and compared in terms of relevance, cost, and performance. Voyage-3-large emerged as the top-performing model, with a wide gap between its performance and that of the second-place group. Voyage-3-lite was found to be a strong option for those looking for high relevancy at a lower cost. Stella, an open-source model, performed well out-of-the-box and can be fine-tuned for better performance. Astra DB provides a complete data API and integrations that make it easier to build production applications with high relevancy and low latency. The results suggest that there is no single "best" model, but rather a range of options depending on specific use cases and requirements.

To improve the retrieval process in a retrieval-augmented generation (RAG) application, consider implementing agentic hybrid search by combining structured metadata with large language model (LLM) decision-making capabilities. This approach enables a smarter and more adaptable system that can handle nuanced queries with greater accuracy. By leveraging the LLM to analyze the query and dynamically select the best retrieval strategy, you can provide several key benefits, including enhanced performance without major overhauls, improved user satisfaction, and increased reliability. With agentic hybrid search, your RAG application can tackle exploratory research, multistep reasoning, and domain-specific tasks while maintaining accuracy, ultimately unlocking its full potential.

The author has built a tool called ResumAI, which uses retrieval-augmented generation (RAG) and generative AI to alleviate the stress of the job hunt by automating the process of improving one's resume for a desired role. The tool utilizes DataStax Langflow, Astra DB, and OpenAI to parse resumes, compare them to existing job descriptions, and provide suggestions for improvement. To build ResumAI, users need to create a serverless vector database in Astra DB, set up an integration with OpenAI, and clone a GitHub repository containing the necessary code. The tool then uses Langflow's AI application builder to create a flow that ingests data from a dataset of LinkedIn job listings, vectorsize the content, and performs a vector similarity search against the uploaded resume. The results are then used to prompt an LLM (Large Language Model) to generate suggestions for improvement. The author has also created a Streamlit app that connects to this Langflow flow, allowing users to input their desired role and upload a resume PDF file to receive personalized feedback on how to improve it.

Bluesky is a social network built on the AT Protocol, an open and decentralized way to build social applications. The AT Protocol allows developers to create custom feeds, clients, and bots. With the advent of generative AI, it's now possible to build smarter chatbots that can have realistic conversations and act independently as agents. Langflow is a drag-and-drop IDE that makes it easier to build AI agents and workflows. To connect clever bots to social media, developers need to create a Bluesky bot powered by Langflow. The application will take control of a Bluesky account, responding whenever it receives a mention or a reply. This can be achieved using the `@skyware/bot` package to interact with the Bluesky API and Langflow to generate responses. To build this application, developers need to set up their environment by creating a new Node.js project, installing dependencies, and configuring the compiler. They also need to create a `.env` file to store account credentials and a `src/index.ts` file to start building the bot code. The bot will listen for "reply" and "mention" events and respond with the author's name and text. Once the bot is ready, developers can define a function to generate content using Langflow and post it in response to the original post. This example bot currently has the personality of a snarky teenager, but developers can experiment with Langflow to create new bots with different personalities.