TimesFM-2.5: A New Standard for Zero-Shot Forecasting

This is seriously a game-changer for anyone who deals with data over time. 🤯 For years, predicting things like future sales, energy demand, or supply chain needs meant building custom, one-off models for every single problem. It was slow, expensive, and a massive headache. Now, imagine having one single, foundational model that you can just point at almost any time-series dataset and get state-of-the-art forecasts. That’s the revolution here. It's like going from a box of specific-sized wrenches to a universal power tool. Google's new TimesFM-2.5 is that tool, and it's an absolute beast: - It's crushing previous benchmarks by up to 25%. - It does this while being twice as efficient (half the size!), making it practical to actually use. - It can look at a much longer history of your data to make smarter predictions. This is massive because it democratizes high-level forecasting. We're moving from building from scratch to fine-tuning a powerful base model. So excited to see what gets built with this! It's already on Hugging Face, so you can go play with it right now - all the links in Yossi's post. Let me know what you think!

Vibe Analytics just dropped plain-language BI. Here's what nobody's talking about: → Plain-language queries finally make BI usable for non-technical users. → Requires reasonably clean data and basic governance. Garbage in, nonsense out. → Article (Oct 13, 2025) positions this for SMBs wanting faster sales, churn, inventory answers. Not gonna lie, this surprised me. Are you already using natural language querying in your workflow? Drop a 🚀 if you want the full breakdown #VibeAnalytics #NoCodeBI #NaturalLanguageQuery #DataGovernance

"Should I use RAG or fine-tune my model?" I get asked this every week. Here's what I learned after building both, multiple times. THE COMMON ADVICE: "RAG for dynamic data, fine-tuning for style." Sounds simple, right? It's not. WHAT ACTUALLY HAPPENED: Built a customer support bot. Tried both approaches. RAG version: Retrieved docs, answered questions, worked in 2 weeks Fine-tuned version: 3 months training, $8K in compute, broke when we updated policies Guess which one shipped? THE REAL DECISION POINTS: After building this 4 times, here's what actually matters: Use RAG when: → Your data changes frequently (weekly/monthly updates) → You need to cite sources and show where answers come from → You want to update knowledge without retraining → You have limited ML expertise Use Fine-tuning when: → You need consistent output format/style that doesn't change → Your domain has specialized terminology that RAG retrieval misses → Response time is critical (fine-tuned = faster inference) → Your knowledge is stable and rarely updates THE REALITY NOBODY MENTIONS: Fine-tuning sounds powerful. But: → Training takes weeks, not days → Every update requires full retraining → You need quality labeled data (expensive to create) → Inference costs are 6x higher than base models RAG sounds simple. But: → Retrieval quality makes or breaks everything → You're maintaining a database + model infrastructure → Bad chunking = bad answers (we rebuilt this 3 times) → Latency from retrieval + generation adds up WHAT WORKS IN PRODUCTION: Most successful implementations I've seen? 80% RAG for facts and data 20% fine-tuning for specific output formats You don't have to choose one. Combine them: → Fine-tune for style and format consistency → RAG for retrieving current information THE HONEST PATTERN: Start with RAG. It's faster to build and easier to maintain. Only fine-tune when: → RAG retrieval consistently fails → You have stable, specialized domain knowledge → You can afford the compute and maintenance costs I've rebuilt systems that used fine-tuning when RAG would work. Never seen it go the other way. BOTTOM LINE: Don't choose based on what sounds impressive. Choose based on: → How often your data changes → Whether you need traceability → What resources you actually have Most problems? RAG solves them cheaper and faster. What's your experience? Have you tried both approaches? #AI #RAG #FineTuning #MachineLearning #LLM #SoftwareEngineering #ProductionAI #ArtificialIntelligence #MLOps #TechLeadership

You can now bring your data quality metadata directly into your favourite AI tools with the SYNQ MCP 🎉 Our data quality MCP lets you use AI clients such as Claude Desktop, OpenAI, or Cursor, and connect to all kinds of data quality related metadata — from lineage and issue history to code commits, schema changes, data profiling, and more. A few of my favourite things I've seen our customers do with it: A DataOps rota handover workflow that summarises open and closed issues for the week in a table, how long each issue has persisted, and who's responsible. It makes handovers between teams much easier. Strategic testing ideas by combining testing philosophies with context to get tailored test suggestions. It's especially helpful when the agent identifies tests that can be removed because a column is a passthrough already tested upstream. Debugging root causes in a third of the time. This one's obvious, but seeing the agent correlate error logs, upstream issues, and recent code changes to pinpoint the exact root cause is just cool every time. https://lnkd.in/dhFyCihX

Your users want to chat with their data. Your engineers don't have 3 months to build it. Ship production-ready AI data chat in 5 minutes with camelAI's embeddable API. No matter how many dashboards you build, your users always have one more question. Another custom report. Another SQL query. Another sprint derailed. Your engineers are drowning in ad-hoc data requests while your users wait days for answers. camelAI breaks this cycle. Their agent writes SQL in real-time, explores your data iteratively, and generates interactive visualizations, all while learning your business terminology and respecting row-level security for multi-tenant architecture. Works with Postgres, MySQL, Snowflake, BigQuery, ClickHouse, and more. They're processing 40,000+ queries monthly across Fortune 500s and startups. Try it for free at camelai.com/api - no sales call required. Congrats on the launch, Illiana Reed, Isabella Reed, and Miguel Salinas! https://lnkd.in/g3jV-m4H

Here's my hot take: if you're still building RAG 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗾𝘂𝗲𝗿𝘆 𝗮𝗴𝗲𝗻𝘁, you're missing out on tons of performance. Query agents are the best thing you can do to easily improve your RAG performance. Why? Because users are messy. In the real world, user interactions with chatbots are unclear, incomplete, and often not what your system expects. Most RAG pipelines assume users will ask perfect questions, but that's rarely the case. Instead: • Users ask vague questions like "tell me about the thing from last month" • They use different terminology than your documents • They ask complex multi-part questions that need decomposition • They don't know how to phrase queries for optimal retrieval 𝗤𝘂𝗲𝗿𝘆 𝗔𝗴𝗲𝗻𝘁𝘀 don't just search, instead, they reason about what the user actually wants. They can: • Expand unclear queries into multiple specific searches • Break down complex questions into manageable parts • Route queries to the right data sources • Reformulate questions for better retrieval But bet you wouldn’t guess that even in ideal situations, they still work wayyy better for retrieval. We just ran comprehensive benchmarks comparing Weaviate's Search Mode (using our Query Agent) against traditional hybrid search across 12 major Information Retrieval benchmarks including BEIR, LoTTe, and BRIGHT. And yea, Search Mode consistently outperformed hybrid search: • +𝟭𝟳% 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 in Success @ 1 (finding the right document first) • +𝟭𝟭% 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 in Recall @ 5 (finding relevant documents in top 5) • Minimum improvement was +5%, maximum was +24% These are industry-standard benchmarks testing everything from scientific papers to customer support queries, and it works way beyond the best traditional methods. 𝗛𝗼𝘄 𝗦𝗲𝗮𝗿𝗰𝗵 𝗠𝗼𝗱𝗲 𝗪𝗼𝗿𝗸𝘀: Under the hood, it applies decades of information retrieval best practices: • Query expansion and decomposition • Schema introspection • Automatic reranking • Dynamic collection routing The agent reasons about your query, understands your schema and data structure, and orchestrates multiple retrieval strategies to get you the best results. Yes, there's a latency trade-off - these additional model inferences take slightly longer. But if your use case prioritizes quality and accuracy over milliseconds, query agents are absolutely the answer. TLDR: The era of naive RAG is over. Users don't ask perfect questions, so why should our systems expect them to? Want to see the full benchmarking results? Check out the detailed analysis: https://lnkd.in/gsh2iYPz

Here's my hot take: if you're still building RAG 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗾𝘂𝗲𝗿𝘆 𝗮𝗴𝗲𝗻𝘁, you're missing out on tons of performance. Query agents are the best thing you can do to easily improve your RAG performance. Why? Because users are messy. In the real world, user interactions with chatbots are unclear, incomplete, and often not what your system expects. Most RAG pipelines assume users will ask perfect questions, but that's rarely the case. Instead: • Users ask vague questions like "tell me about the thing from last month" • They use different terminology than your documents • They ask complex multi-part questions that need decomposition • They don't know how to phrase queries for optimal retrieval 𝗤𝘂𝗲𝗿𝘆 𝗔𝗴𝗲𝗻𝘁𝘀 don't just search, instead, they reason about what the user actually wants. They can: • Expand unclear queries into multiple specific searches • Break down complex questions into manageable parts • Route queries to the right data sources • Reformulate questions for better retrieval But bet you wouldn’t guess that even in ideal situations, they still work wayyy better for retrieval. We just ran comprehensive benchmarks comparing Weaviate's Search Mode (using our Query Agent) against traditional hybrid search across 12 major Information Retrieval benchmarks including BEIR, LoTTe, and BRIGHT. And yea, Search Mode consistently outperformed hybrid search: • +𝟭𝟳% 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 in Success @ 1 (finding the right document first) • +𝟭𝟭% 𝗶𝗺𝗽𝗿𝗼𝘃𝗲𝗺𝗲𝗻𝘁 in Recall @ 5 (finding relevant documents in top 5) • Minimum improvement was +5%, maximum was +24% These are industry-standard benchmarks testing everything from scientific papers to customer support queries, and it works way beyond the best traditional methods. 𝗛𝗼𝘄 𝗦𝗲𝗮𝗿𝗰𝗵 𝗠𝗼𝗱𝗲 𝗪𝗼𝗿𝗸𝘀: Under the hood, it applies decades of information retrieval best practices: • Query expansion and decomposition • Schema introspection • Automatic reranking • Dynamic collection routing The agent reasons about your query, understands your schema and data structure, and orchestrates multiple retrieval strategies to get you the best results. Yes, there's a latency trade-off - these additional model inferences take slightly longer. But if your use case prioritizes quality and accuracy over milliseconds, query agents are absolutely the answer. TLDR: The era of naive RAG is over. Users don't ask perfect questions, so why should our systems expect them to? Want to see the full benchmarking results? Check out the detailed analysis: https://lnkd.in/gsh2iYPz

𝗪𝗲 𝗴𝗼𝘁 𝗱𝗿𝘂𝗻𝗸 𝗼𝗻 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀. 🍹 The party was fun... but now comes the hangover. Last week at the Great Data Debate during Big Data LDN, I was honored to join an incredible panel with Jeremiah Stone, Vivienne Wei, Mary Osborne, and David Kalmuk, moderated by Mike Ferguson. One question sparked a lively discussion that I wish we had even more time for: 𝗪𝗵𝗲𝗿𝗲 𝗮𝗿𝗲 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀 𝗵𝗲𝗮𝗱𝗲𝗱? As teams balance the operational burden of yet another database to keep in sync against rapidly improving AI models, this question becomes increasingly critical. 𝗠𝘆 𝗲𝘅𝗽𝗮𝗻𝗱𝗲𝗱 𝗽𝗲𝗿𝘀𝗽𝗲𝗰𝘁𝗶𝘃𝗲: The pendulum is swinging back. While embeddings have their place, classic lexical search is making a comeback, especially for code assistants. We're seeing strong results from:  • Hybrid approaches (think BM25 + rerankers like Anthropic's contextual retrieval)  • Sometimes skipping embeddings entirely (leveraging expanded context windows and/or agent-driven traditional search) Supporting evidence:  • Anthropic's Contextual Retrieval: https://lnkd.in/gG_uWpzP  • Why Cursor is About to Ditch Vector Search: https://lnkd.in/gqpD2RHs  • Why Cline Doesn't Index Your Codebase: https://lnkd.in/gVwBs5Z2 𝗧𝘄𝗼 𝗺𝗮𝗷𝗼𝗿 𝘁𝗮𝗶𝗹𝘄𝗶𝗻𝗱𝘀 𝗮𝗿𝗲 𝗿𝗲𝘀𝗵𝗮𝗽𝗶𝗻𝗴 𝘁𝗵𝗲 𝗹𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲:  1. Context windows continue growing rapidly  2. Agentic workflows (bots performing multiple targeted searches) deliver surprisingly strong results Meanwhile, PostgreSQL's (among others) native vector support challenges the need for separate systems, eliminating multi-vendor operational overhead. 𝗠𝘆 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹 𝘀𝘂𝗴𝗴𝗲𝘀𝘁𝗶𝗼𝗻𝘀:  • 𝘚𝘵𝘢𝘳𝘵 𝘴𝘪𝘮𝘱𝘭𝘦: Use lexical search or direct context injection  • 𝘐𝘧 𝘦𝘮𝘣𝘦𝘥𝘥𝘪𝘯𝘨𝘴 𝘢𝘳𝘦 𝘯𝘦𝘤𝘦𝘴𝘴𝘢𝘳𝘺: Consolidate with a single system first (PostgreSQL + pgvector)  • 𝘜𝘴𝘦 𝘢 𝘥𝘦𝘥𝘪𝘤𝘢𝘵𝘦𝘥 𝘷𝘦𝘤𝘵𝘰𝘳 𝘋𝘉: only as a last resort when you've proven the need I'm curious: how does this match with your experience? Where have vector databases accelerated or hindered your team's productivity?

We crossed an important milestone last week. We rolled out AI-powered features in SQL Prompt at no extra cost for all subscribers. Our approach is simple: bring practical, trusted AI directly into the tools database professionals already use, and do it in a way that supports - not replaces - their expertise. The new features, including natural language-to-SQL, clear explanations of queries, and intelligent index analysis, are designed to make everyday work faster, clearer and more effective. But most importantly, they’re opt-in and transparent, putting the user in full control. This is the next step in Redgate’s mission to give database professionals the confidence to do their best work: pairing innovation with reliability, and always building for real-world use. If you’re a SQL Prompt or SQL Toolbelt Essentials subscriber, you’ll see these features available, and you can read more about them here:  https://lnkd.in/eRD8denx