Weaviate

Single vs. Multi-Agent Systems: Which is right for your project? Learn all the key patterns in multi-agent systems and download our free ebook to get you started! We cover: - Single vs multi-agent systems - Patterns in multi-agent systems - 6 examples of agentic architectures - ... and much more! Download https://lnkd.in/djNFdcHP

Code > Concept: AGNTCY's Internet of Agents is now on GitHub Just weeks after launching AGNTCY, we're shipping code for the Internet of Agents, tackling the multi-agent software lifecycle from Discover—>Deploy: — DISCOVER: Open Agent Schema Framework and Agent Directory – to define agent capabilities and be able to catalog and search for agents at scale. — COMPOSE: Agent Connect Protocol, IO Mapper Agent and API Bridge Agent (as part of a multi-agent software toolkit) - to allow for agent-agent communication at the sematic and syntactic layers — DEPLOY: Agent Gateway and Workflow Server – for messaging and runtime on current cloud native state of the art — EVALUATE: in the works, but will tackle Observability, Evaluation, and Security frameworks especially for multi-agent systems. We are also welcoming 4 new members to the AGNTCY family: MongoDB, Traceloop, Agency,?and Weaviate, alongside our current members Outshift by Cisco, LangChain, Galileo??, LlamaIndex, Glean Finally, we are introducing AGNTCY Supporter – a new role in the collective we’ve created for companies who believe in the mission of an open agent ecosystem. We're building the foundations that lets AI agents talk to each other regardless of who built them, where they run, or what framework they're using. Two big calls to action for you: 1.?Read Papi Menon blog to understand how these components fit together and what they actually do 2. Try it out. Fork the repo, open issues, submit PRs. We need your participation to build an Internet of Agents that works for everyone, not just walled gardens. Blog & GitHub Links in comments ?? #AGNTCY #InternetOfAgents #OpenSource

Stop trying to find internal company info like this.. This no-code AI solution takes 5 minutes to build. Sometimes it pays to work smarter, not harder - I built this company knowledge Q&A system with zero coding, using StackAI's Platform. These customizable, no-code solutions provide huge benefit to companies wanting to set up AI workflows without needing developer resources - or teams who are looking to leverage AI in minutes, not weeks. Check it this webinar to learn more! https://lnkd.in/eEjiFesq Awesome company, awesome product - using Weaviate's vector database under the hood ??

Ever curious about what "Q4_K_M" meant or what quantization is in the world of LLMs and vector embeddings? Marcin Antas is a Weaviate Core Engineer and in his video, he talks about the power of quantization in AI?deployment. Here are some key points he talks to in his video: - LLMs store?knowledge in floating-point parameters (16-bit/8-bit) - Quantization converts these?to smaller 4-bit formats - GGUF format with LLAMA.cpp?is now the industry standard - Feed-forward layers (60% of parameters) contain most?of the model's knowledge Embedding?models can be optimized using ONNX for specific CPU architectures. Weaviate uses three quantization approaches: ? Product quantization (up to 90% compression) ? Binary quantization (32x?reduction, less accurate) ? Scalar?quantization (balance of precision and size) https://lnkd.in/gGwT3AER

Looking for all the essential best practices for Weaviate? We now have a comprehensive, evolving guide ? Whether you're just starting with Weaviate or looking to optimize your existing implementation, bookmark this page and check back regularly. Here are some example best practices you should follow: 1?? ???????????? ???????????? ? Explicitly define your schema instead of relying on auto-schema functionality ? Avoid cross-references when possible - they're not vectorized and can slow down queries ? Consider flattening your data structure for better vector representation 2?? ???????? ???????????? ???????? ?????????????????? ? Always use batch imports for datasets larger than 10 objects ? Batch imports reduce network overhead and enable more efficient vectorization ? For maximum throughput, consider using Go or Java client libraries 3?? ???????????????? ???????????????????? Memory Requirements Guide: ? 6GB for 1M vectors (1024 dimensions) ? 1.5GB for 1M vectors (256 dimensions) ? 2GB for 1M vectors with quantization enabled 4?? ?????????????????????? ???????????????????????? ? Consider vector quantization to reduce memory usage while maintaining performance ? For single-tenant collections under high load, disable lazy loading ? Use asynchronous client APIs in async environments for better performance 5?? ??????????-?????????????? ???????? ? Offload infrequently accessed tenants to cold storage to reduce costs ? Keep lazy loading enabled for multi-tenant deployments ? Use dynamic indexes for better resource management in multi-tenant setups Pro Tip: Keep your Weaviate instance and client libraries up-to-date to benefit from the latest features and improvements Want to learn more? Check out our detailed best practices page: https://lnkd.in/dB5C43Kv

“RAG is dead" - Social media, whenever context windows expand. Is it though? "I've watched 'RAG is dead' trend 5 times this year. Meanwhile, Agentic RAG is silently revolutionizing how AI accesses information." Every few months, someone proclaims RAG's death because "now we have 100k/million token context windows!" Yet somehow, RAG keeps showing up to work on Monday morning, completely unaware of its own demise. Why RAG refuses to die: ??♂? ? LLMs don't know what happened after training cutoff ? They can't access your private data unless shown ? They hallucinate when uncertain (making up legal cases is apparently a hobby) ? Constant retraining is expensive But regular RAG was just the beginning. The real evolution is Agentic RAG. Traditional RAG: "Here's some relevant context I found. Hope it helps!" Agentic RAG: "Let me break down your complex question, search multiple sources, evaluate what I find, and give you a comprehensive answer while explaining my reasoning." What makes it "agentic"? 1. ???????????? ?? ????????: Breaking complex queries into manageable sub-queries 2. ?????????????????? ???????? ??????????: Using vector search, APIs, and other external tools 3. ???????????????????? ?????? ??????????????????: Assessing results and adjusting until satisfied Agentic RAG can: ? Decompose complex queries into simpler sub-queries ? Determine when it needs more information from you ? Evaluate the relevance and accuracy of retrieved information ? Reformulate queries when results aren't satisfactory ? Adapt its search strategy based on initial findings So the next time someone tells you "RAG is dead because of long context windows," just smile and say: "That's weird... my agentic RAG system just planned a complex research strategy, executed it across multiple data sources, evaluated the results, and delivered a perfect answer. For a dead technology, it sure is productive." Learn about all Agentic Workflows in this free ebook: https://lnkd.in/dpxRtjxf

?? Help us help other AI builders ?? What do you want to see in a guide for Embedding ???? model selection & evaluation? Here is a ?? draft ?? outline for the upcoming Weaviate Academy ?? module on the topic. https://lnkd.in/edczJESz It's centred around the concept of iterative improvements on these key ideas. - Identify your needs - Compile candidate model list - Perform detailed evaluation / comparison - Roll out & maintain ?? I would love to know what you all think. - What's missing? What do you want to know? - What do you wish you could tell your co-workers, or yourself when you started your AI building journey? Let me know here in comments, or comment on the page directly. #BuildingInPublic #AI

Is MCP just function calling with a fancy name? No - But I get the confusion.? Let me break it down. Here’s how I would explain the difference ?? · When you want to provide new tools to an LLM, you use it via function calling. But that tool has to actually be running somewhere. · ?????? ???????????????????????? ?????? ???? ???????? ?????????? ?????????????? and live, as well as how LLMs or other AI applications actually ???????????? these running tools. · An MCP server and a Tool form a perfect duo, where the MCP server provides an easy way for you to make that tool accessible to an LLM. · The key here is that ? you do not have to re-invent the wheel for each tool you need ?. So… Do you need an MCP server? Maybe, maybe not.. Why “maybe”? The answer is “depends on who you’re talking to”. ???? AI engineers who've been building LLM tools for a while might have custom solutions already in place. But why re-invent the wheel? ?? MCP is like REST for LLMs - a simple, repetitive protocol that lets any developer make tools accessible to LLMs without worrying about how to serve and expose them. Many MCP servers exist already, including Weaviate's MCP server which lets you access Weaviate collections as a knowledge base or memory store. ?? Create your Weaviate account to get started: https://lnkd.in/en6CuvQC ???? Check out the Weaviate MCP server: https://lnkd.in/e2CKUxA8

What makes up an agent? At first it wasn’t clear, so we did some digging. Turns out it’s not that complicated. AI agents are made up of just a few core components. We identified these 3 as the most important: ???Reasoning (via LLMs) Agents use specialized LLMs to plan, execute, and refine tasks iteratively by breaking down complex problems (planning) and adjusting their future actions based on outcomes (reflection). ????Tools AI agents extend their capabilities beyond their training data by using external tools such as web search, APIs, and databases, enabling two-way interaction with their environment—both retrieving real-time information and making changes to external systems. ???Memory Memory enables AI agents to retain context and learn from past interactions, distinguishing agentic workflows from purely LLM-driven ones. Short-term memory helps track immediate context for decision-making, while long-term memory stores knowledge across sessions for personalization and continuous improvement. Other elements for agents include: ???Structured prompt defining role and task ??♀??Human feedback loops for refinement and decision-making ???Permissions that determine level of autonomy Want to better understand agents? We’ve got heaps of resources: ???Agentic Architectures e-book: https://lnkd.in/dqKxCcq6 ???Agentic Workflows blog post: https://lnkd.in/dtZkcDGg

???Just announced: Morningstar won a 2025 CIO 100 award for their Intelligence Engine Platform! Huge congrats to Benjamin Barrett, Aisis Julian, and their team for developing such an innovative project driving real business value for their organization. ?????Check out their story here: https://lnkd.in/gQW57H_f ??Learn more about the CIO 100 and the other winners here: https://lnkd.in/g7Pherzn