Vijay Morampudi’s Post

Large Language Models (LLMs) on Structured Data Sets: A Synopsis of Current Challenges and Research Directions The integration of Large Language Models (LLMs) with structured data sets has become a focal point of discussion among senior technical leaders responsible for data architecture in various companies. This summary encapsulates the essence of recent conversations and research efforts aimed at understanding and optimizing the use of LLMs in the context of structured data. The primary application of LLMs in structured data environments is to interpret human queries and translate them into SQL or other query languages. This translation process is currently seen as inefficient, as it not only demands substantial computational resources but also does not fully leverage the capabilities of LLMs to address the underlying problem statements. The convolution arises from the mismatch between the unstructured nature of LLMs and the rigid, predefined schemas of structured data. Research in this domain is actively seeking to bridge the gap between the fluidity of LLMs and the rigidity of structured data. The goal is to enhance the user experience by creating more intuitive and efficient ways for LLMs to interact with and extract information from structured databases. This endeavor is not without its challenges, as the computational intensity of LLMs poses scalability issues. A striking comparison reveals that a typical LLM request consumes approximately 17 times more power than a standard Google search query. This disparity highlights the need for innovation in making LLMs more energy-efficient and cost-effective, especially when dealing with structured data. The sustainability and scalability of LLMs in data architecture are contingent upon advancements that can reconcile their high computational demands with the economic and environmental costs. One such research effort is the work by Wang et al. (2021), which proposes a novel framework for integrating LLMs with relational databases. Their approach involves a pre-processing step that transforms structured data into a format more suitable for LLMs, thereby reducing the computational overhead. Another significant contribution is the research by Zhang and Choi (2020), which focuses on optimizing the interaction between LLMs and structured data by introducing an intermediary layer that can effectively translate natural language queries into database operations. The ongoing research in this field is a testament to the potential of LLMs to revolutionize data architecture. The quest for a more harmonious integration of LLMs with structured data sets is not merely a convolution but a glimpse into the future of data management. The innovation that successfully mitigates the current limitations will undoubtedly pave the way for a new multi-billion dollar market, offering unprecedented opportunities for businesses and researchers alike. #LLMs #DataArchitecture #StructuredData #Innovation #Technology #Research

"Day 6: Query Construction - Bridging the Gap Between Language and Databases!" Hey everyone! Today, we’re exploring Query Construction—an essential step in the RAG flow where your query is reshaped to match the format of different databases. 🛠️✨ In this step, the query is transformed from natural language into a form that specific data systems can understand. This ensures that the LLM can retrieve data accurately from various sources. Two key methods used in this process are Text-to-SQL and Text-to-Cypher: 1. Text-to-SQL: This transformation converts natural language queries into SQL queries. For example, if you ask, "What are the top-selling products?" the system transforms this into an SQL query like SELECT * FROM products WHERE sales > X. This works great for structured databases like relational databases. 2. Text-to-Cypher: Cypher is a query language used for graph databases (like Neo4j). When the LLM needs to search a graph database, it transforms the query into a Cypher query. For instance, "Show connections between company X and person Y" might transform into a Cypher query like MATCH (a:Company)-[:RELATED_TO]->(b:Person) RETURN a, b. This is useful for databases that focus on relationships between data points. Query Construction acts like a translator between the language we speak and the languages databases understand, ensuring that no matter how complex the database is, the LLM can communicate with it effectively. 🔄📊 learn more about here https://lnkd.in/e4pAuuUt Tomorrow, we’ll explore the Retrieval step—where we start pulling data based on these transformed queries! Stay tuned! 😊✨ #rag #genai #llm #ai #data #AI #Chatbots #TechInnovation #MachineLearning #FutureOfWork

Knowledge Graphs provide 300% Higher Accuracy for LLM Responses in Enterprises "A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases" #seo #ai #llm https://lnkd.in/egbQb9Dd

🔔 Democratizing Data: How Natural Language Interfaces are Empowering End Users to Interact with Databases 🔴 Hermes: A Text-to-SQL Solution at Swiggy 💠 Hermes is a generative AI-based workflow developed by Swiggy to facilitate data accessibility for its teams. The tool allows users to input natural language questions and receive corresponding SQL queries and results directly in Slack. This streamlines the data access process, enabling faster and more efficient decision-making. 🔶 The Need for Hermes 💠Many business and product decisions require specific numbers and quantities that are often locked away in databases, accessible only to those with SQL knowledge. 💠Traditional methods of data access, like searching dashboards or requesting data from analysts, can be time-consuming and inefficient. 💠Hermes democratizes data access, making it faster and easier for everyone to get the information they need. 🔶Key Features and Benefits 💠Natural Language Interface: Users can ask questions in plain English, eliminating the need for SQL expertise. 💠Instant Results: Hermes automatically generates SQL queries and executes them, delivering results directly in Slack within minutes. ( a data cleansing layer via Lamda is also there) 💠Improved Data Accessibility: Empowers users across different roles to access and analyze data independently. 💠Enhanced Decision-Making: Enables faster, data-driven decisions by providing quick access to critical information. 💠Increased Efficiency: Streamlines the data querying process, saving time and effort for users. 🔶Technology Behind Hermes 💠Generative AI: Leverages the power of large language models (LLMs) like GPT 3.5 and 4.0 to generate SQL queries. 💠Knowledge Base and RAG: Incorporates Swiggy-specific context through a knowledge base and Retrieval-Augmented Generation (RAG) techniques. 💠Data Catalog: Integrates with Swiggy's in-house data catalog, Lumos, for metadata management. 💠Cloud Computing: Utilizes AWS Lambda for middleware and Databricks for job creation and query execution. 🔴I have seen a similar capability embedded within Oracle Autonomous Database (ADB). For enterprises using Oracle Autonomous Database and seeking to implement a robust and user-friendly Text-to-SQL solution, leverage the power of the built-in "speak human" AI ( Select AI, a dbms package) feature in conjunction with a middleware layer like AWS Lambda / OCI functions for enhanced functionality and seamless integration. ( reference : https://lnkd.in/dCbgFFMs). 🔴 By enabling interaction with databases using natural language, these tools break down the barriers of technical expertise, allowing individuals across different roles to leverage data for informed decision-making and problem-solving. As technology continues to advance, we can expect even more intuitive and user-friendly solutions that further democratize data access and empower individuals to harness the power of information for better outcomes.

Table-Augmented Generation (TAG), an advanced paradigm to address the limitations of Text2SQL and Retrieval-Augmented Generation (RAG) when querying databases using natural language. Problem Statement: - Text2SQL only works for queries expressible in SQL, limiting its use to questions directly translatable into relational algebra. - RAG handles simple lookups but struggles with more complex queries that involve reasoning or aggregation across multiple data rows or external knowledge. - Many real-world questions require reasoning that goes beyond what SQL can express (e.g., sentiment analysis or summarizing trends). Table-Augmented Generation (TAG): TAG consists of three main steps: -- Query Synthesis: Converts a user’s natural language question into an executable database query (e.g., SQL). -- Query Execution: Executes the query on a database to retrieve the relevant data. -- Answer Generation: Uses the language model to interpret the retrieved data and generate a final natural language response. TAG integrates semantic reasoning from LMs with the data aggregation and filtering capabilities of databases. Why TAG is Needed: TAG bridges the gap between Text2SQL (good for structured queries) and RAG (good for simple lookups), covering a wide range of user queries involving both structured and unstructured data. Evaluation: Handwritten TAG pipelines significantly outperform these methods, achieving accuracy improvements of 20% to 65%. Limitations of Text2SQL and RAG: - Text2SQL methods translate natural language into SQL, but are limited to queries expressible as SQL commands. They cannot handle semantic tasks like summarization or classification. - RAG works for simple lookups but struggles with large-scale or complex queries that require aggregation, reasoning, or handling large volumes of data. TAG Architecture: - TAG combines the strengths of both SQL (efficient data processing and querying) and LMs (handling natural language reasoning). It is composed of: - Query Synthesis: Translates the natural language question into a query for the database. For instance, if the user asks for reviews of the highest-grossing romance movie, TAG generates the appropriate SQL query. - Query Execution: Executes the query on the database to retrieve relevant data (e.g., the title, genre, revenue, and reviews of movies). - Answer Generation: Uses an LM to process the retrieved data and generate a human-readable answer. For example, TAG might summarize reviews using the LM’s semantic capabilities. Advantages of TAG: - TAG enables more complex interactions between language models and databases than previously possible with Text2SQL or RAG. Capability: - Answer queries that require contextual knowledge (e.g., understanding what constitutes a “classic” movie). - Perform reasoning over multiple rows of data (e.g., summarizing customer reviews or computing trends). - Combine exact computation from databases with language understanding from LMs. #text2sql

Don’t Build Your Future on Specialized Vector Databases https://lnkd.in/gfTxqiYr With the rise of AI, vector databases have gained significant attention due to their ability to efficiently store, manage and retrieve large-scale, high-dimensional data. This capability is crucial for AI and generative AI (GenAI) applications that deal with unstructured data such as text, images and videos. The main logic behind a vector database is to provide similarity search capabilities, rather than keyword search, as traditional databases provide. This concept has been widely adopted to boost the performance of large language models (LLMs), particularly following the release of ChatGPT. The biggest issue with LLMs is that they require substantial resources, time and data for fine-tuning. Which makes it very difficult to keep them updated. This is why when you query LLMs about recent events, they often provide answers that are factually incorrect, nonsensical or disconnected from the input prompt, leading to “hallucinations.” One solution is retrieval-augmented generation (RAG), which augments an LLM by integrating up-to-date information retrieved from an external knowledge base. Specialized vector databases are designed to handle vectorized data efficiently and provide robust semantic search capabilities. These databases are optimized for storing and retrieving high-dimensional vectors, which are very important for making similarity searches. The speed and efficiency of vector databases have made them an integral part of RAG systems. The hype around vector databases has led many people to suggest that traditional databases might be replaced by vector databases. Instead of storing data in traditional (SQL or NoSQL) databases, could you store an organization’s entire data set in a vector database and retrieve it using natural language instead of writing manual queries? But vector databases don’t function like traditional databases. As Qdrant CTO Andrey Vasnetsov wrote, “the majority of vector databases are not databases in this sense. It is more accurate to call them search engines.” This is because their main purpose is to provide optimized search functionalities, and they are not designed to support basic features like keyword search or SQL queries. Limitations of Specialized Vector Databases As use cases grew and people focused on the scalability of their applications, the limitations of vector databases became more visible. Developers soon realized they still need the features of a full-text search engine along with vector search. For example, filtering search results based on specific criteria is very difficult with vector databases. These databases also lack direct matches for exact phrases, which are crucial for many tasks. Limited Support for Complex Queries Complex queries often involve multiple conditions, joins and aggregations, making them challenging for specialized vector databases. These databases provide limited support ...

Previously our approach on text to sql is not producing SQL up to the mark, For better results we are using concept call Retrieval Augmented Generation. #Day 6 🚀 What is Retrieval-Augmented Generation (RAG)? 🚀 As the world of AI rapidly evolves, a powerful method called Retrieval-Augmented Generation (RAG) is gaining traction. If you're into AI, natural language processing, or even database management, RAG is a game-changer you should know about! 🌟 🔍 What is RAG? In a nutshell, RAG is a hybrid AI approach that combines the best of two worlds: 1. Retrieval: The model first retrieves relevant information from a database, document store, or external source based on the user's input. 2. Generation: Then, it uses this retrieved information to generate a more accurate and context-aware response. 💡 Why Does It Matter? Instead of generating responses purely from pre-trained knowledge, RAG integrates real-time, relevant data into the output. For example, if you ask an AI model a question about your company’s sales data, it can first fetch the specific data you need and then generate a detailed SQL query to analyze that data. Here’s why it’s so powerful: - Accuracy: With RAG, you’re not relying solely on the AI’s general knowledge; it can access external information to ensure its responses are based on the latest or most relevant data. - Context-Awareness: RAG enables AI to understand the specific context of your query, resulting in more meaningful and actionable outputs. - Scalability: Whether it's fetching documents, SQL data, or research papers, RAG ensures your AI assistant scales with your growing knowledge base or data. 💬 How It’s Used: - Automated SQL Generation: Imagine asking a natural language question like "Show me sales data for the last quarter." RAG can retrieve the relevant data schema and then generate an SQL query tailored to your specific database. - Personalized Assistance: It can fetch personal or organizational documents to give you personalized responses in customer support, legal advice, or research. 🔧 Practical Example: When building a tool to convert natural language into SQL queries, RAG can first retrieve the correct data schema or relevant information from your database, making the SQL generation more accurate and contextual. It brings the intelligence of retrieval-based search and the creativity of generation together in one powerful model. In Summary: RAG boosts the intelligence and relevance of AI by empowering it to look up data before generating responses. Whether you're working with SQL databases or any text-based system, RAG can make your workflows smarter and faster. #RAG #AI #MachineLearning #NLP #DataScience #Automation #SQL #sadakpramodh

🚀 Unlock the Power of Generative LLMs for Text to SQL Generation! 💻✨ Ask questions in natural language and get your SQL queries generated! 🔍 Text to SQL is revolutionizing how we interact with databases by converting natural language queries into executable SQL. However, generating accurate SQL for your specific schema can be quite challenging, even for advanced large language models (LLMs). 🎉 But don't worry, there's no need to reinvent the wheel! Some fantastic tools are already out there to help: 🔧 Enter Vanna: An open-source Python retrieval-augmented generation (RAG) framework for SQL generation. Vanna works in two simple steps: Create embeddings with the data definition language statements (DDLs) and sample SQLs for your schema. Have a look at their whitepaper to understand how RAG has really transformed the accuracy is Text to SQL Generation, https://lnkd.in/gpBjpWmt 💡 This RAG implementation is one way, but you could also leverage Foundational models specifically trained for SQL generation: ✨ NSQL: A new family of open-source large foundation models designed specifically for SQL generation tasks. NSQL is database-agnostic and accelerates the development of customized enterprise Foundation models in analytics workflows. It’s the first pretrained text-to-SQL model that surpasses all existing open-source models by up to 6 points in code execution accuracy. 🔍 Learn more about how NSQL tackles the challenges faced by existing models and elevates your analytics workflows! https://lnkd.in/gEB-ZpJ7 #AI #TextToSQL #NLP #LLMs #NaturalLanguageProcessing #RAG #NSQL #Vanna #Analytics

Tired of writing complex SQL queries? Select #AI unlocks the power of Oracle Autonomous Database through natural language conversations. Ask questions, get insights in multiple languages. Learn more: https://lnkd.in/ehpJMqhc