Weaviate

What if your AI agents could search 10,000 internal documents in milliseconds with full HIPAA compliance? 𝗦𝘁𝗮𝗰𝗸𝗔𝗜 + 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲 just made it possible without writing code. StackAI handles the orchestration letting teams build AI agents visually with drag-and-drop workflows that connect directly to enterprise systems like Salesforce, SharePoint, and ServiceNow. Weaviate powers the retrieval layer - indexing internal documents, policies, and knowledge bases with hybrid search and metadata filtering, then retrieving the right context in milliseconds. Together, they enable RAG systems that are: • Secure by design (SOC 2, HIPAA, GDPR compliant) • Built with granular permissions and access controls • Fast and accurate at enterprise scale • Fully auditable and traceable This turns RAG from a prototype into production-grade AI infrastructure that enterprises can actually trust. The full technical guide covers advanced RAG techniques, real workflow architectures, and best practices for safety and monitoring. Download it here: https://lnkd.in/eayWKYiZ

Most AI chatbots fail the moment you ask two things at once. And there's a simple reason why. Here's a typical conversation: 𝗬𝗼𝘂: "I want to order a pepperoni pizza. Does it contain gluten? Also, how long will delivery take?" 𝗕𝗼𝘁: "Our pepperoni pizza is made with wheat-based dough and contains gluten." 𝗬𝗼𝘂: "...and the delivery time?" 𝗕𝗼𝘁: "I'm sorry, I can only help with menu and ingredient questions. Let me transfer you to our delivery team." Sound familiar? This happens because most AI agents route your message to ONE specialist. The Menu Agent got your question. It answered its part. Everything else? Not its problem. Now multiply this across every multi-part question: • "Cancel my subscription and refund last month's charge" • "Book the 3 pm flight and add extra baggage" • "Show me my recent orders and update my address" Users think in complete tasks. Systems fragment them. Elysia (open-source from Weaviate) solves this - it handles natural, free-flowing conversations where users ask multiple things at once without breaking apart. Wrote a deep dive into how it works. Thanks to Danny Williams, Victoria Slocum, and the folks at Weaviate for building and open-sourcing this. Github link to the project: https://lnkd.in/dYucGM73 Blogpost link in comments.

𝗦𝗺𝗮𝗹𝗹 𝗖𝗵𝗮𝗻𝗴𝗲, 𝗕𝗶𝗴 𝗜𝗺𝗽𝗮𝗰𝘁 (𝗗𝗮𝘆 𝟯/𝟱): 𝟭𝟮𝘅 𝗥𝗲𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗼𝗳 𝗜𝗻𝘁𝗲𝗿-𝘇𝗼𝗻𝗮𝗹 𝘁𝗿𝗮𝗳𝗳𝗶𝗰 ⚡️📉 Another day, another significant win. Today it's not a performance update, but a traffic optimization. Learn how we reduced traffic by 12x on a large customer's cluster with 1,700+ updates per second. When you do an update in Weaviate, Weaviate makes a decision whether only the object is updated ("metadata-only update") or whether also the vector is updated. This was primarily an optimization to reduce CPU overhead because the cost of updating the vector index is significantly higher than that of "only" updating the object store and its adjacent indexes (bitmap, inverted, block-max, etc.) But little did we know, this had a hidden side effect that became especially visible in larger clusters. While the coordinating node (i.e. the one that initially receives the request) already determined that the vector is unchanged, it still passed the vector along to the other nodes. The other nodes treated this as an idempotent update and internally turned the vector update into a no-op, just like the coordinator did. But at that point, the data was already sent over the wire. So all we had to do is not transfer the vector from the coordinator to the other replicas when it's clear that the vector did not change. While this is an incredibly simple optimization, the effect was huge for this customer's cluster. They have a fairly high update rate of 1,700+ RPS (24/7), so this simple optimization dropped the traffic from a 120MB/s average down to just 10MB/s. 🤯 That's a whopping 11TB of interzonal traffic saved every month for—what is essentially—a single-line code change. (Yes, I put those em dashes there, this not AI-generated 🤣 ) What's your biggest "tiny change, massive impact" success story?

What separates a ChatGPT wrapper from a production-grade agentic system? 𝗧𝗵𝗲 𝗮𝗻𝘀𝘄𝗲𝗿: Two CEOs are revealing their playbook for reliable AI agents at scale. We're teaming up with CrewAI for a workshop on building agentic workflows that actually scale. Sign up here: https://lnkd.in/emzj9fkZ 𝗖𝗿𝗲𝘄𝗔𝗜 is a multi-agent framework for orchestrating autonomous AI agents. Instead of relying on a single agent to handle complex tasks, CrewAI lets you create teams of specialized agents - each with specific roles, tools, and goals. These agents collaborate, delegate, and refine each other's outputs to solve problems that would overwhelm a single agent. CrewAI integrates with Weaviate through the 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲𝗩𝗲𝗰𝘁𝗼𝗿𝗦𝗲𝗮𝗿𝗰𝗵 tool. This enables your agents to run semantic search queries over documents stored in your Weaviate cluster, turning retrieved knowledge into context-aware responses. Your agents can access both web search (via Serper API) and your vector database - combining fresh context with domain-specific knowledge. 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂'𝗹𝗹 𝗹𝗲𝗮𝗿𝗻: • What it takes to design and build agents with Crew • How Weaviate is shaping the future of retrieval • Why external tools are essential for building reliable agentic applications The workshop features João (Joe) Moura (CEO, CrewAI) and Bob van Luijt (CEO, Weaviate) - so you'll be learning directly from the people building these tools. Whether you're exploring multi-agent systems or looking to move your prototypes into production, this workshop covers the orchestration, tooling, and database architecture you need to build agents that scale.

𝗦𝗺𝗮𝗹𝗹 𝗖𝗵𝗮𝗻𝗴𝗲, 𝗕𝗶𝗴 𝗜𝗺𝗽𝗮𝗰𝘁 (𝗗𝗮𝘆 𝟮/𝟱): 𝗙𝗮𝘀𝘁𝗲𝗿 𝗿𝗲-𝘀𝗰𝗼𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗰𝗼𝗺𝗽𝗿𝗲𝘀𝘀𝗲𝗱 𝗛𝗡𝗦𝗪 (𝗣𝗤/𝗦𝗤/𝗥𝗤/𝗕𝗤) ⚙️ ⚡ Any time you have compression, you have rescoring. Today's hidden improvement is a 25%+ speed-up from faster rescoring from better re-use of resources. First of all: why rescore? What is it and why is it needed? Vector Compression can introduce slight accuracy losses because the compressed distance only approximates the exact distance. This can easily be overcome by doing a slight over-fetch and re-score with the original vectors (from disk). Such an over-fetch is already quite normal in HNSW through the ef value (which is typically set higher than the limit), so re-scoring fits really well: All you need to do is get the full-precision vectors for your final candidate set before trimming the results and re-sorting them according to the real distances. But this means you have more work on the hot path, so it needs to be efficient. That's exactly what we improved here. How did we do it? It's all about re-use. ♻️ Re-using memory is faster than allocating new memory. But the biggest speed-up came from reusing the consistent view from Weaviate's LSM storage engine: Rather than isolating each vector retrieval, the new implementation now opens a single view and uses it for rescoring all vectors. This increased throughput by about 25% (at the same recall) for all compression methods: Product (PQ), Scalar (SQ), and Rotational Quantization (RQ) But even better yet; it compounds with the Day 1 improvements from yesterday because PQ also uses rescoring. With both changes combined, you get more than a 100% speed-up of PQ queries between just two Weaviate patch versions! Thanks for joining for yet another day. I promise tomorrow's hidden improvement is going to be about something other than performance improvements for compressed queries. What's your favorite technique to speed up performance? Branch avoidance? Re-use? Something else?

C# and Java developers: the wait is over. Weaviate’s C# and Java clients are now GA! Weaviate now supports five client languages: - Python - JavaScript/TypeScript - Go - C# - Java With the addition of C# and Java clients, you now have even more flexibility to build with Weaviate in the language you love. Happy building! 💚

Small Change, Big Impact: Day 1/5: PQ Speed-up 📈 ⚙️ Product Quantization just got ~60% faster on average between Weaviate v1.34.7 and v1.34.8. ⚡ This week, I'm doing a mini series on "Hidden Improvements" in Weaviate. Stuff that's too cool not to share, but might otherwise get lost somewhere in a release note. I kid you not, the idea for this came because I watched a random Matthias Wandel video (https://lnkd.in/dtwva5DC) over the holidays. In the video, he makes a few changes to decade-old code. That got me thinking: I'd love to do a hardcore optimization like in the early days. Before the break I talked with our Field-CTO about a specific cluster where PQ calculations showed up more than usual in their CPU profiles, so I thought why not give this a shot. The technique I used is probably as old as coding itself: I removed branches from the hot path. Our PQ implementation uses a Look-up-Table that was filled on-demand (lazily). The simple check if we already had an entry completely stalled CPUs. So, now what we're doing instead, is precalcuate it all. You may think an exhaustive precalculation of all possible options is too expensive to do on the hot path, but it's cheaper than all those branches the compiler/runtime can't predict. Probably one of the most basic optimization techniques out there, but makes a huge difference! That's it for today, join me again tomorrow for day 2 of hidden improvements. Also if you like code optimizations, AI/vibe coding, (or my random YouTube highlights), give me a follow!

Vector databases finally feel native in C#! After months of development and feedback, we're excited to release the 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲 𝗖# 𝗰𝗹𝗶𝗲𝗻𝘁 - bringing the full power of our vector database to the .NET ecosystem. This isn't just an API wrapper. We built a modern, type-safe client that feels native to C# developers: ✅ 𝗥𝗔𝗚-𝗿𝗲𝗮𝗱𝘆 - built-in support for vector search + generative AI ✅ 𝗖𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻-𝗰𝗲𝗻𝘁𝗿𝗶𝗰 𝗔𝗣𝗜 - intuitive operations grouped around collections ✅ 𝗦𝘁𝗿𝗼𝗻𝗴𝗹𝘆-𝘁𝘆𝗽𝗲𝗱 - full IntelliSense support with generic types ✅ 𝗗𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝘆 𝗶𝗻𝗷𝗲𝗰𝘁𝗶𝗼𝗻 - seamless integration with modern .NET apps Getting started is simple: dotnet add package 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲.𝗖𝗹𝗶𝗲𝗻𝘁 --version 1.0.0 Whether you're building semantic search, RAG applications, or AI-powered features, you can now do it all in C# with a developer experience we're really proud of. Huge thanks to Michelangelo P. for leading this effort! 💚 Read the full announcement: https://lnkd.in/et27pfN5