SQL Server Performance Survival Kit at Techorama Belgium

Most RAG systems retrieve from one source. Ours retrieves from three — and decides which one to use on the fly. I co-authored and presented a research paper titled "Adaptive Hybrid Agentic Retrieval-Augmented Generation using LangGraph for Intelligent Knowledge Systems" at the International Conference on Artificial Intelligence, Machine Learning and Big Data Engineering (ICAIMLBDE), Nagercoil — March 30, 2026. Here's the problem we solved: Enterprise knowledge doesn't live in one place. It's split across unstructured documents, structured databases, and relational knowledge graphs. Traditional RAG systems only know how to search vector databases — so the moment a query involves structured data or entity relationships, they either retrieve the wrong thing or retrieve nothing useful at all. What we built: A multi-agent RAG architecture that routes every query to the right retrieval strategy automatically. → A Supervisor Agent built with LangGraph analyzes each query and decides which retrieval path to activate → Vector Database Agent handles semantic, unstructured queries using cosine similarity → SQL Database Agent handles structured queries requiring exact matches and filters → Knowledge Graph Agent handles relationship-based reasoning between entities → A Context Aggregation module merges results from all active agents before passing to the LLM → A Response Validation Agent verifies the final answer is grounded in retrieved knowledge The key insight: not all enterprise queries are the same shape. A query like "what are the leave policies?" needs semantic search. A query like "show me all employees who joined after 2023 with rating above 4" needs SQL. A query like "how is Project A related to Employee B?" needs a knowledge graph. A static pipeline treats all three identically — and fails on two of them. An agentic pipeline routes intelligently. The difference in answer quality is significant. If your RAG system only speaks to a vector database — it's only seeing part of your enterprise knowledge. #GenerativeAI #RAG #LangGraph #AgenticAI #MultiAgent #LargeLanguageModels #MachineLearning #ResearchPaper #ICAIMLBDE #KnowledgeGraphs

Building a Graph RAG System without a Graph Database Graph RAG without a graph database. Sounds contradictory, but the "graph" in Graph RAG is about connecting entities across documents to answer multi-hop questions. Most solutions reach for Neo4j or similar: a second database to deploy, a new query language to learn, two systems to keep in sync. A different approach: store entities, relations, and passages in 3 Milvus collections linked by IDs. The graph structure lives inside the vector database as cross-references. Traversal is just a series of ID lookups. No Cypher, no separate infra. The retrieval pipeline is 4 steps: 1. Seed retrieval — vector search for relevant entities 2. Subgraph expansion — follow ID links to discover bridge entities 3. LLM rerank — one call to filter noise 4. Answer generation — one call to produce the final answer Two LLM calls total. Iterative approaches (IRCoT, Agentic RAG) need 3-10+. Results across 3 multi-hop QA benchmarks: - 87.8% avg Recall@5 (beats HippoRAG 2 on average) - ~60% lower API costs vs iterative methods - 2-3x faster end-to-end Runs locally with Milvus Lite. No servers needed to get started. By Cheney Zhang. Link in comments #GraphRAG #Milvus #VectorDatabase #MultiHopReasoning #OpenSource -- 📩 The Year of the Graph Spring 2026 newsletter issue is out! Beyond Context Graphs: How Ontology, Semantics, and Knowledge Graphs Define Context 👇 https://lnkd.in/d7u3Yh-B All things #KnowledgeGraph, #GraphDB, Graph #Analytics / #DataScience / #AI and #SemTech. Subscribe and follow to be in the know. Reach out if you'd like to be featured

Tech Saturday Today's Technology Demystified session from the Department of Artificial Intelligence and Data Science focused on Vector Databases and Embedding Search Vector Databases: While traditional databases (like PostgreSQL) manage rows and columns, vector databases are optimized for storing and querying multi-dimensional vector data. • Fast Retrieval: They use Approximate Nearest Neighbor (ANN) algorithms to find similar vectors in milliseconds, even across billions of data points. • Key Features: Beyond indexing, they provide CRUD operations, metadata filtering (e.g., "find similar items under $50"), and scalability. The Embedding Search Works as follows: 1. Indexing: Source data is embedded by a model and stored in the vector database. 2. Querying: When a user searches, the query itself is converted into a vector using the same embedding model. 3. Similarity Search: The database calculates the mathematical distance (e.g., Cosine Similarity or Euclidean distance) between the query vector and stored vectors. 4. Ranking: The "nearest neighbours" (most similar items) are returned as the search results. A few Key Use Cases of Vector Databases and Embedding Search • Retrieval-Augmented Generation (RAG): Providing relevant context from a private knowledge base to Large Language Models (LLMs) to reduce hallucinations. • Recommendation Engines: Suggesting products or movies based on their semantic similarity to what a user has previously liked. • Multimodal Search: Searching for an image using a text description (e.g., "sunset at the beach") or vice versa. • Anomaly Detection: Identifying outliers in financial transactions or network logs that fall far from "normal" behavior cluster The session highlighted ✓. What is Vector Databases and Embedding Search? ✓. Why does it matter? ✓. How it works? ✓. Real world Applications ✓. The future of Vector Databases and Embedding Search ✓. Job roles emerging with Vector Databases and Embedding Search #TechSaturday #SRMTRPEC #DepartmentofArtificialIntelligenceandDataScience #VectorDatabasesandEmbeddingSearch

Not everything needs to start big. I built a simple RAG-based “Chat with PDF” app to understand how embeddings, retrieval, and LLMs actually work together in practice. Yes — it’s basic. But it unlocked something important. As someone with a strong background in Data engineering & SQL performance tuning and database optimization, I’m now taking this further — building toward a system that can: • Analyze database DDL • Ask intelligent follow-up questions • Generate queries for real metrics (table size, index usage, etc.) • Suggest performance optimizations based on actual data Moving from static document Q&A → to intelligent, system-aware optimization. This is just the starting point. The real work begins now. #DataEngineering #AI #RAG #LLM #DatabaseOptimization #AIDataEngineering

⚡ How Vector Databases Power RAG (and Why They're Different) Yesterday I talked about why traditional databases struggle with RAG — they search by keywords, but RAG needs search by meaning. Vector databases flip the model entirely. Here's how ⬇️ ━━━━━━━━━━━━━━━━━━━━ 𝗧𝗵𝗲 𝗕𝗶𝗴 𝗦𝗵𝗶𝗳𝘁 Instead of storing text with inverted indexes, vector databases store 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 — dense numerical representations that capture semantic meaning. Your query → converted to a vector of 384 numbers Your documents → also vectors in the same space Documents about "authentication sequence", "user session initialization", and "OAuth token exchange" all cluster nearby — even with zero keyword overlap. The DB finds closest vectors using cosine similarity. 𝗬𝗼𝘂'𝗿𝗲 𝘀𝗲𝗮𝗿𝗰𝗵𝗶𝗻𝗴 𝗯𝘆 𝗺𝗲𝗮𝗻𝗶𝗻𝗴, 𝗻𝗼𝘁 𝗯𝘆 𝘄𝗼𝗿𝗱𝘀. ━━━━━━━━━━━━━━━━━━━━ 𝗪𝗵𝘆 𝗧𝗵𝗶𝘀 𝗠𝗮𝘁𝘁𝗲𝗿𝘀 𝗳𝗼𝗿 𝗥𝗔𝗚 3 reasons from real experience: ① 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗾𝘂𝗲𝗿𝗶𝗲𝘀 𝗷𝘂𝘀𝘁 𝘄𝗼𝗿𝗸 "How does the widget system resolve pages?" → Finds WidgetContextPageService without mentioning the class name ② 𝗖𝗼𝗱𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗶𝘀 𝗺𝘂𝗹𝘁𝗶-𝗹𝗮𝘆𝗲𝗿𝗲𝗱 Embeddings understand that takeUntilDestroyed() relates to "subscription cleanup" — even if no comment says so ③ 𝗧𝘄𝗼-𝘀𝘁𝗮𝗴𝗲 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 Find files by summary → then chunks within those files Traditional DB needs complex joins across both levels ━━━━━━━━━━━━━━━━━━━━ 𝗪𝗵𝗮𝘁 𝗩𝗲𝗰𝘁𝗼𝗿 𝗗𝗕𝘀 𝗚𝗶𝘃𝗲 𝗬𝗼𝘂 ✦ Native ANN indexing (HNSW) ✦ First-class embedding storage ✦ Metadata filtering + vectors ✦ Simpler mental model — data = vectors, queries = vectors ChromaDB, Pinecone, Weaviate, Qdrant — all purpose-built for this. But does that mean traditional DBs are useless in RAG? Not at all. 🔜 Tomorrow: Real-world RAG architecture combining both. ━━━━━━━━━━━━━━━━━━━━ ♻️ Repost to help your network 💬 Which vector DB are you using? #VectorDatabase #RAG #AI #MachineLearning #LLM #ChromaDB #Embeddings #ENTAIN

Recently reviewed a database schema generated using AI for a PDF analytics system. At first glance, it looked solid: 1. Clean structure 2. Logical relationships 3. Good starting design But when analyzing deeper from a production perspective, a few important gaps appeared: 1. Logging tables without growth strategy 2. No aggregation layer for analytics 3. Session tracking without structure 4. Redundant data storage 5. Missing query-aligned indexes None of these are “mistakes” , they are simply things that only show up when thinking about: • Scale • Performance • Real-world usage AI is doing a great job at generating schemas. But designing systems that hold under load and evolve over time still needs deeper engineering thinking. That’s where experience complements AI. #DatabaseDesign #SystemDesign #AI #BackendEngineering #MySQL #MuraliCodes

"Yugabyte stands out as a provider that blends quick rollout of innovative features with stronger support for more traditional database capabilities."💬 Read this ISG Software Research Analyst Perspective from ISG Director of Research Matt Aslett, outlining how YugabyteDB meets AI needs for operational data platforms.💡 https://gag.gl/QMC3Wk #distributedsql #RAG #MCPServer #Postgrescompatible #ai #GenAI #database #cloudnative #Observability #yugabytedb #opensource #Vector

𝗔𝗿𝗲 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀 𝗯𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗼𝗽𝘁𝗶𝗼𝗻𝗮𝗹 𝗶𝗻 𝗥𝗔𝗚 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲𝘀? 🤔 Building a RAG application with Vector DBs was just the beginning. ** 𝑾𝒆𝒍𝒄𝒐𝒎𝒆 𝒕𝒐 𝒕𝒉𝒆 𝒏𝒆𝒘 𝒆𝒓𝒂 𝒐𝒇 𝑽𝒆𝒄𝒕𝒐𝒓𝒍𝒆𝒔𝒔 𝑹𝑨𝑮 ** For the last couple of years, Retrieval-Augmented Generation (RAG) has largely revolved around 𝐞𝐦𝐛𝐞𝐝𝐝𝐢𝐧𝐠𝐬 + 𝐯𝐞𝐜𝐭𝐨𝐫 𝐝𝐚𝐭𝐚𝐛𝐚𝐬𝐞𝐬. While powerful, this approach comes with its own challenges: ❌ High infrastructure cost (embedding + storage + retrieval) ❌ Latency overhead due to similarity search ❌ Loss of semantic precision in certain edge cases ❌ Continuous re-indexing pipelines ⏲️ Now we’re seeing a shift towards "𝐕𝐞𝐜𝐭𝐨𝐫𝐥𝐞𝐬𝐬 𝐑𝐀𝐆" Instead of relying heavily on vector similarity, newer approaches leverage: ✅ Structured retrieval (SQL / graph-based querying) ✅ Metadata-driven indexing ✅ LLM-native retrieval (reasoning over raw or lightly indexed data) ✅ Hybrid search (symbolic + semantic without strict vector dependency) 🎯 The focus should be moving from 𝐚𝐩𝐩𝐫𝐨𝐱𝐢𝐦𝐚𝐭𝐞 𝐬𝐢𝐦𝐢𝐥𝐚𝐫𝐢𝐭𝐲 𝐭𝐨 𝐜𝐨𝐧𝐭𝐞𝐱𝐭-𝐚𝐰𝐚𝐫𝐞 𝐩𝐫𝐞𝐜𝐢𝐬𝐢𝐨𝐧. Where does this really matter? 1️⃣ 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝗞𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 Traditional RAG struggles with highly structured data like financial reports, compliance documents, or relational datasets. Vectorless RAG can directly query structured sources, ensuring accuracy, traceability, and better governance. 2️⃣ 𝗥𝗲𝗮𝗹-𝘁𝗶𝗺𝗲 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 In use cases like supply chain optimization or fraud detection, latency and freshness are critical. Avoiding heavy embedding pipelines enables faster, near real-time retrieval and response generation. 👨🏼💻👩🏼💻𝗪𝗵𝗮𝘁 𝗱𝗼𝗲𝘀 𝘁𝗵𝗶𝘀 𝗺𝗲𝗮𝗻 𝗳𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀/𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀? 👉 Rethink whether every problem needs embeddings 👉 Design retrieval as a data problem first, not just an AI problem 👉 Combine LLM reasoning with deterministic data access patterns 👉 Focus on explainability and auditability, not just relevance scores Lastly, we’re not replacing vector databases, we’re simply evolving beyond dependency on them. 🎬 The future of RAG isn’t just smarter models. 𝑰𝒕’𝒔 𝒔𝒎𝒂𝒓𝒕𝒆𝒓 𝒓𝒆𝒕𝒓𝒊𝒆𝒗𝒂𝒍. ⁉️ Vectors or no vectors, what’s your bet for the next 12 months? #AI #GenAI #RAG #VectorDatabases #DataEngineering #LLM #ArtificialIntelligence #KnowledgeSystems #AIArchitecture #FutureOfAI

🚀 Stop Choosing Between Vector and Graph Databases—Use Both. In 2026, the "Vector vs. Graph" debate has matured. As architects, we’re no longer asking which is better; we’re asking which relationship we need to capture. 🔹 Vector Databases (The "Vibe" Check) Vectors excel at Semantic Similarity. They are the workhorse of RAG because they handle messy, unstructured data (PDFs, chats, logs) without a pre-defined schema. Best for: "Find documents that sound like this query." The Weakness: They treat data as isolated chunks. They struggle with "multi-hop" reasoning (e.g., "Find the manager of the person who approved Project X"). 🔹 Graph Databases (The "Fact" Check) Graphs excel at Explicit Relationships. They use nodes and edges to provide absolute precision and explainability. Best for: "How are these two entities connected?" The Weakness: High friction. You need a schema, and "semantic" nuances (like synonyms) aren't inherent—you have to define them. 💡 The Architect’s Pivot: GraphRAG The real winner in enterprise AI right now is GraphRAG. By layering a Knowledge Graph over a Vector store, we solve the two biggest RAG headaches: Hallucinations: Graphs provide "ground truth" constraints. Context Loss: Graphs preserve connections across document boundaries that vector chunks often miss. 🏗️ My Rule of Thumb: Use Vector when the data is unstructured and the query is "exploratory." Use Graph when the relationships are the data (Fraud, Supply Chain, IAM). Use Both when you need your LLM to reason like an expert. #DataArchitecture #VectorDatabase #GraphDatabase #GenerativeAI #RAG #GraphRAG #TechLeadership #Wipro #CTOInsights

"Yugabyte stands out as a provider that blends rapid delivery of innovative features with stronger support for traditional database capabilities." Check out this ISG Software Research Analyst Perspective from ISG Director of Research Matt Aslett, which explains how YugabyteDB meets AI needs for operational data platforms. https://gag.gl/QMC3Wk #distributedsql #RAG #MCPServer #Postgrescompatible #ai #GenAI #database #cloudnative #Observability #yugabytedb #opensource #Vector