SQL remains essential in data science landscape

Is SQL still important in the age of LLMs? Alex makes a strong case that the answer is "yes," particularly for RAG applications. Throughout the RAG pipeline—from data preprocessing to retrieval to response evaluation—SQL databases can provide significant advantages for organizing metadata, applying filters, tracking data provenance, and supporting analytics. When a system combines both SQL and vector capabilities, it becomes even more powerful: a unified platform that stores structured metadata alongside vector embeddings, enabling efficient hybrid retrieval by combining semantic search with traditional filtering. Curious which SQL-plus-vector systems work best for your use cases?

🚨 𝗪𝗲 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝗲𝗱 𝗮 𝗰𝗿𝗶𝘁𝗶𝗰𝗮𝗹 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 (𝗖𝗩𝗦𝗦 𝟗.𝟖) 𝗶𝗻 𝗮 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝘂𝘀𝗲𝗱 𝗯𝘆 𝟖𝟕% 𝗼𝗳 𝗙𝗼𝗿𝘁𝘂𝗻𝗲 𝟏𝟎𝟎𝟎 🚨 Today, at Cyera Research, we uncovered a critical vulnerability (CVE-2025–64712, CVSS 9.8) in 𝘂𝗻𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱.𝗶𝗼 library - an ETL library used by most of Fortune 1000 companies as an enabler for their AI transformation. The vulnerability we've found is an Arbitrary File Write via Path Traversal bug that allows threat actors to execute code on the machine running the unstructured library and ultimately takeover the machine. Link to the full blog post in the first comment 👇

Do yourself a favor and go read this research. This vuln is very cool. Like “drop what you’re doing and open the link” cool.

Yesterday, Cyera research found a critical vulnerability (CVSS 9.8) in unstructured.io library, an ETL library used by 87% of Fortune 1000 companies. "...The vulnerability we found is a classic path traversal bug that leads to arbitrary file write. In simple words, it allows a threat actor to write a file of any type, with any content, to anywhere on the file system of the machine running the unstructured library..." Read the full article here: https://lnkd.in/eiFP9_Yx

The latest edition of the Data Science Briefing is now out! This week’s theme is moving from “cool model” to “reliable system.” We highlight an end-to-end RLHF-from-scratch deep dive, practical guidance on writing higher-quality code with AI (without sacrificing correctness), and why production-grade agents need operational memory—not just chat history. We also include a solid Bayesian on-ramp to causal discovery, plus fresh academic reads on where LLM “reasoning” breaks down and what better RL objectives might look like. Check it out! 👉 https://lnkd.in/eAGv8cEf

The rapid enhancement of AI's ability to write computer code has led to considerable angst in the data science and computer science communities. This obviously leads to the critical question of how these groups can redefine their unique value add. I have previously argued (https://lnkd.in/g-sbra7i) for the statistics profession to accept and "own" the challenge of data quality, and recently worked with Willis Jensen to make the same suggestion to the data science community. (https://lnkd.in/ga7-nTjx) In my view, the challenge of data quality is a "jump ball" for whoever has the will to accept it. Enjoy!

After years in the relational world, I’m officially moving into Graph Tech. 🚀 Relational databases are great, but Graph is the "secret sauce" for the world's most complex data challenges. It’s the difference between seeing dots and seeing the whole picture. Current status: Reached "Oxymoron Level 1" - this is not a thing: Contextual Knowledge Graph. It's like Cold Ice. We’re building a Composite AI Investigation Platform to turn fragmented data into real-world intelligence. The future isn't in rows - it's in the connections. #Graph+AI-Platform #KnowledgeGraph #CompositeAI #DecisionIntelligence #InvestigationIntelligence DataWalk

Data science may appear to be all about intricate models and AI wizardry, but it’s actually a collaborative, business-oriented, and iterative process. Solomiia Leno, a Data Scientist at SoftServe, sheds light on several misunderstandings and provides insight into the true nature of this role in action. Learn more about what it’s like at SoftServe: https://sftsrv.com/UJrMIA #DataScience #SoftServe

Decision Trees and Why Ensembles Are Necessary A single Decision Tree is like that one friend who is incredibly logical but also incredibly biased. They have a 'Rule' for everything. 'If the temperature is > 30 and the day is Sunday, then everyone will buy ice cream.' It sounds smart, but it’s fragile. If it snows on a Sunday (a rare event), the tree will still confidently tell you to buy ice cream. This is called 'Overfitting.' To build a production system, you don't need one 'Logical Friend.' You need a 'Jury.' --- THE ENSEMBLE: THE WISDOM OF THE CROWD --- An Ensemble method (like Random Forest or XGBoost) is just a collection of many decision trees. - Random Forest (Bagging): You train 100 trees on different random subsets of your data. Then you let them vote. If 90 trees say 'Spam' and 10 say 'Not Spam,' the answer is Spam. The 'crowd' averages out the individual biases of each tree. - Boosting (XGBoost): You train one tree. It makes mistakes. Then you train a second tree specifically to *fix* the mistakes of the first one. You repeat this until your error is nearly zero. --- WHY TREES STILL BEAT DEEP LEARNING (FOR TABULAR DATA) --- If your data lives in a spreadsheet (CSV, SQL, Excel), Ensembles of Trees almost always beat Neural Networks. 1. They handle 'Missing Data' effortlessly. 2. They don't care about the 'Scale' of your numbers (you don't need to normalize). 3. They are significantly faster to train and deploy. --- THE SARCASTIC BOTTOM LINE --- Stop trying to build a 'Neural Network' for your tabular business data. It’s like using a telescope to read a book. XGBoost or Random Forest is the workhorse of the modern enterprise. It’s not 'sexy,' but it’s what actually powers the models that detect fraud, predict churn, and determine insurance rates across the world. Are you a 'Deep Learning snob' or do you acknowledge the power of the mighty Random Forest? 👇 #MachineLearning #RandomForest #XGBoost #DataScience #AIEngineering #SoftwareEngineering #EnsembleLearning #TabularData