Milvus, created by Zilliz

You’ve optimized everything — indexes, filters, batching. But searches are STILL slow during certain times? Query nodes share resources with background compaction, upserts, and migrations. Each steals CPU and memory from your queries. 🔍 𝐒𝐢𝐠𝐧𝐚𝐥𝐬 𝐭𝐨 𝐰𝐚𝐭𝐜𝐡 𝐟𝐨𝐫: ✓ CPU spikes during background tasks ✓ Disk I/O saturation ✓ Slow searches after restarts ✓ Many small, unindexed segments 𝐇𝐨𝐰 𝐭𝐨 𝐟𝐢𝐱 𝐢𝐭: • 𝐑𝐞𝐬𝐜𝐡𝐞𝐝𝐮𝐥𝐞 𝐜𝐨𝐦𝐩𝐚𝐜𝐭𝐢𝐨𝐧 to off-peak hours • 𝐁𝐚𝐭𝐜𝐡 𝐮𝐩𝐬𝐞𝐫𝐭𝐬 so compaction keeps up • 𝐑𝐞𝐥𝐞𝐚𝐬𝐞 𝐮𝐧𝐮𝐬𝐞𝐝 𝐜𝐨𝐥𝐥𝐞𝐜𝐭𝐢𝐨𝐧𝐬 to free memory • 𝐖𝐚𝐫𝐦 𝐮𝐩 𝐜𝐚𝐜𝐡𝐞𝐬 after restarts • 𝐔𝐩𝐠𝐫𝐚𝐝𝐞 𝐫𝐞𝐠𝐮𝐥𝐚𝐫𝐥𝐲 for bug fixes • 𝐃𝐞𝐝𝐢𝐜𝐚𝐭𝐞 𝐫𝐞𝐬𝐨𝐮𝐫𝐜𝐞𝐬 to query nodes Think about ecosystems, not just a query engine. Learn how to debug in Milvus in detail: https://lnkd.in/gK9QUjGb ——— 👉 Follow Milvus, created by Zilliz, for everything related to unstructured data!

🚗 Volvo Cars 𝐛𝐮𝐢𝐥𝐭 𝐭𝐫𝐮𝐬𝐭𝐰𝐨𝐫𝐭𝐡𝐲 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞 𝐀𝐈 𝐬𝐲𝐬𝐭𝐞𝐦𝐬 𝐰𝐢𝐭𝐡 𝐌𝐢𝐥𝐯𝐮𝐬 — transforming document chaos into reliable retrieval. As one of the world’s most respected automakers, Volvo Cars needed 𝐞𝐧𝐭𝐞𝐫𝐩𝐫𝐢𝐬𝐞-𝐠𝐫𝐚𝐝𝐞 𝐑𝐀𝐆 𝐭𝐨 𝐝𝐞𝐥𝐢𝐯𝐞𝐫 𝐟𝐚𝐬𝐭, 𝐞𝐱𝐩𝐥𝐚𝐢𝐧𝐚𝐛𝐥𝐞 𝐚𝐧𝐬𝐰𝐞𝐫𝐬 𝐚𝐜𝐫𝐨𝐬𝐬 𝐏𝐃𝐅𝐬, 𝐬𝐥𝐢𝐝𝐞𝐬, 𝐬𝐩𝐫𝐞𝐚𝐝𝐬𝐡𝐞𝐞𝐭𝐬, 𝐚𝐧𝐝 𝐦𝐨𝐫𝐞. However, their existing cloud vendor’s tool fell short — hallucinations, vague citations, and an unsustainable pricing model made scaling difficult. ⚙️ 𝐖𝐡𝐲 𝐌𝐢𝐥𝐯𝐮𝐬 Milvus stood out for its real-world usability — practical, notebook-level examples guided the team through handling complex documents, choosing embeddings, and building production-ready RAG pipelines. As Hanlian L. Product Owner & BI Expert at Volvo Cars, noted: “Milvus’s real advantage was how easy and friendly it made things to understand and execute.” 📊 𝐑𝐞𝐬𝐮𝐥𝐭𝐬: 1️⃣ Up to 90% lower costs compared to the previous AI Search service 2️⃣ Transparent, verifiable search results that stakeholders trust 3️⃣ Outperformed the previous cloud vendor’s solution in side-by-side demos — more reliable, faster, and easier to validate We’re proud to help Volvo Cars build an enterprise-grade RAG system that’s both technically robust and organizationally trusted — reflecting the future of enterprise AI: scalable, transparent, and built on trust. 🔗 https://lnkd.in/gY7pMBd7 GitHub: https://lnkd.in/gpcK-eK —— Follow Milvus, created by Zilliz for practical tips and real-world use cases on vector search at scale. #BuiltwithMilvus

Your filters are optimized. Batching is perfect. But searches are still slow? Choosing the wrong vector index can tank your performance — in-memory indexes are fast but memory-hungry, while on-disk indexes save memory at the cost of speed. 🔍 𝗦𝗶𝗴𝗻𝗮𝗹𝘀 𝘁𝗼 𝘄𝗮𝘁𝗰𝗵 𝗳𝗼𝗿: ✓ High Vector Search Latency in 𝗤𝘂𝗲𝗿𝘆 Node metrics ✓ Disk I/O saturation with DiskANN or 𝗠𝗠𝗔𝗣 ✓ Slower queries right after restart (cold 𝗰𝗮𝗰𝗵𝗲) 𝗛𝗼𝘄 𝘁𝗼 𝗳𝗶𝘅 𝗶𝘁: 𝗠𝗮𝘁𝗰𝗵 𝗶𝗻𝗱𝗲𝘅 𝘁𝗼 𝘄𝗼𝗿𝗸𝗹𝗼𝗮𝗱 • 𝗛𝗡𝗦𝗪 – In-memory, high recall, lowest latency. • 𝗜𝗩𝗙 – Flexible trade-offs between recall and speed. • 𝗗𝗶𝘀𝗸𝗔𝗡𝗡 – Billion-scale datasets. Needs fast SSD bandwidth. • 𝗠𝗜𝗡𝗛𝗔𝗦𝗛_𝗟𝗦𝗛 – Binary vectors (Milvus 2.6+) with MHJACCARD metric. 𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝘁𝘂𝗻𝗶𝗻𝗴: • Enable 𝗠𝗠𝗔𝗣 to balance latency vs memory • Tune search parameters for your recall needs • Warm up frequently accessed segments after restarts The “best” index depends on your data size, memory budget, and latency requirements. No one-size-fits-all. Learn more: https://lnkd.in/gK9QUjGb Next: Background Noise (when infrastructure fights against you) ——— 👉 Follow Milvus, created by Zilliz for everything related to unstructured data!

LangChain isn’t LangGraph. CrewAI isn’t AutogenAI. Make isn’t n8n. Treating them the same? You'll never build effective AI agents. Here's what each one actually does: 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧: Chains together LLMs, data, and external tools 𝐋𝐚𝐧𝐠𝐆𝐫𝐚𝐩𝐡: Build complex agent workflows with memory and control flow 𝐂𝐫𝐞𝐰𝐀𝐈: Assigns specialized roles and sequences agent collaboration 𝐀𝐮𝐭𝐨𝐠𝐞𝐧𝐀𝐈:  Enables autonomous agent conversations and collaborations 𝐌𝐚𝐤𝐞: Visual no-code platform for automating app integrations 𝐧8𝐧: Open-source workflow automation with code flexibility ——— 👉 Follow Milvus, created by Zilliz, for everything related to unstructured data!

We’re excited to announce that our 𝗠𝗶𝗹𝘃𝘂𝘀 𝗖𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 𝗠𝗲𝗲𝘁𝘂𝗽 will be part of #OpenSourceAIWeek — a weeklong celebration of #OpenSource innovation in #GenAI & ML! 🎉 Spots are filling up fast! Join us next week on 𝗧𝘂𝗲𝘀𝗱𝗮𝘆, 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝟮𝟭 @𝟱:𝟯𝟬 𝗣𝗠 for an exciting evening learning about: 🔥 Jiang Chen diving into RaBitQ, JSON Index, Schema Evolution, What's New in Milvus 2.6 🔎 Sharing on Lessons Learned: Building Great Search Experience in Production ⚡️ Lightning Talk on RAG development at SAP Save your seat before it’s gone 👉 https://luma.com/zoym6vcy Check out Open Source AI Week 👉 https://lnkd.in/gPFZdYQG

Your got efficient vector index, QPS isn't too high, but filtered search are still slow. Why? 𝐓𝐰𝐨 𝐜𝐨𝐦𝐦𝐨𝐧 𝐩𝐫𝐨𝐛𝐥𝐞𝐦𝐬:  Missing the scalar index on the filtered field? That would force a full table scan so filtered search gets slow. Did you use strong consistency level? That will wait until all nodes synchronize before querying. How to 𝐟𝐢𝐱 𝐢𝐭? 𝐈𝐧𝐝𝐞𝐱 𝐭𝐡𝐞 𝐅𝐢𝐞𝐥𝐝 𝐘𝐨𝐮 𝐅𝐢𝐥𝐭𝐞𝐫: No indexes = full scans. 𝐃𝐞𝐟𝐢𝐧𝐞 𝐈𝐧𝐝𝐞𝐱𝐞𝐬 𝐟𝐨𝐫 𝐉𝐒𝐎𝐍 𝐚𝐧𝐝 𝐃𝐲𝐧𝐚𝐦𝐢𝐜 𝐒𝐜𝐡𝐞𝐦𝐚: For JSON fields, Milvus 2.6 introduced path and flat index. 𝐂𝐡𝐨𝐨𝐬𝐞 𝐭𝐡𝐞 𝐑𝐢𝐠𝐡𝐭 𝐂𝐨𝐧𝐬𝐢𝐬𝐭𝐞𝐧𝐜𝐲: Avoid Strong consistency level unless absolutely necessary → Choose Bounded/Eventually when you can. 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐞 𝐅𝐢𝐥𝐭𝐞𝐫 𝐄𝐱𝐩𝐫𝐞𝐬𝐬𝐢𝐨𝐧𝐬: BAD: tag == "A" OR tag == "B" OR tag == "C" → GOOD: tag IN ["A", "B", "C", "D"] What's next: We will dive into choosing the right vector index (HNSW vs DiskANN vs IVF). Stay tuned! ——— 👉 Follow Milvus, created by Zilliz, for everything related to unstructured data!

𝗘𝘃𝗲𝗿 𝗻𝗼𝘁𝗶𝗰𝗲𝗱 𝗔𝗟𝗟 𝘆𝗼𝘂𝗿 𝗾𝘂𝗲𝗿𝗶𝗲𝘀 𝘀𝘂𝗱𝗱𝗲𝗻𝗹𝘆 𝘀𝗹𝗼𝘄 𝗱𝗼𝘄𝗻 𝗮𝘁 𝗼𝗻𝗰𝗲? 𝗧𝗵𝗮𝘁'𝘀 𝘄𝗼𝗿𝗸𝗹𝗼𝗮𝗱 𝗼𝘃𝗲𝗿𝗹𝗼𝗮𝗱. Large NQ (number of 𝗾𝘂𝗲𝗿𝘆 data per request) or high QPS monopolizes query-node resources. Other requests queue up waiting. So how do you fix it? 𝗕𝗮𝘁𝗰𝗵 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁𝗹𝘆: Break requests into 100–500 vectors each, not 10,000 at once. 𝗦𝗰𝗮𝗹𝗲 𝗛𝗼𝗿𝗶𝘇𝗼𝗻𝘁𝗮𝗹𝗹𝘆: Scale out 𝗾𝘂𝗲𝗿𝘆 nodes if high concurrency is sustained. 𝗠𝗼𝗻𝗶𝘁𝗼𝗿 𝗤𝘂𝗲𝘂𝗲 𝗠𝗲𝘁𝗿𝗶𝗰𝘀: Watch in-queue latency. If it's climbing, you need capacity. Next: Inefficient Filtering (and why missing indexes kill performance). 𝗦𝘁𝗮𝘆 𝘁𝘂𝗻𝗲𝗱! ——— 👉 Follow Milvus, created by Zilliz, for everything related to unstructured data!

🧬🚀 𝗙𝗿𝗼𝗺 𝗽𝗿𝗼𝘁𝗲𝗶𝗻 𝗽𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻 𝘁𝗼 𝗰𝗿𝗼𝘀𝘀-𝗺𝗼𝗱𝗮𝗹 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆, 𝗔𝗜 𝗶𝘀 𝗿𝗲𝘄𝗿𝗶𝘁𝗶𝗻𝗴 𝗯𝗶𝗼𝗹𝗼𝗴𝘆 — BioMap chose Milvus to fuel this revolution! As a leading life sciences AI company, 𝗕𝗶𝗼𝗠𝗮𝗽 𝗶𝘀 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗯𝗶𝗼𝗹𝗼𝗴𝗶𝗰𝗮𝗹 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹𝘀 that accelerate drug discovery and medical research. However, the challenges were clear — 𝘀𝗹𝗼𝘄 𝗽𝗿𝗼𝘁𝗲𝗶𝗻 𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝘀, 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗱𝗮𝘁𝗮, and the struggle to balance speed with accuracy across diverse workloads. 🔍 To succeed, they needed infrastructure capable of searching 𝗯𝗶𝗹𝗹𝗶𝗼𝗻𝘀 𝗼𝗳 𝘀𝗲𝗾𝘂𝗲𝗻𝗰𝗲𝘀 𝗮𝗻𝗱 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀 𝗶𝗻 𝗿𝗲𝗮𝗹 𝘁𝗶𝗺𝗲. After testing tools like Faiss, BioMap chose 𝗠𝗶𝗹𝘃𝘂𝘀 as their engine of discovery: 🔓 Open-source flexibility 🧪 Production-ready stability 🔗 Comprehensive features ⚡ Performance at scale With Milvus at its core, BioMap now powers three critical capabilities for biological AI: 🔹 𝗔𝗜 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆 𝗔𝘀𝘀𝗶𝘀𝘁𝗮𝗻𝘁 — a RAG system that delivers 𝘀𝘂𝗯-𝘀𝗲𝗰𝗼𝗻𝗱 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 across scientific literature and biological databases. 🔹 𝗣𝗿𝗼𝘁𝗲𝗶𝗻 𝗦𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲 𝗣𝗿𝗲𝗱𝗶𝗰𝘁𝗶𝗼𝗻 — scaling to 𝟱𝟬𝗕+ 𝘀𝗲𝗾𝘂𝗲𝗻𝗰𝗲𝘀 and running 𝟮𝟮𝘅 𝗳𝗮𝘀𝘁𝗲𝗿 than traditional MSA. 🔹 𝗖𝗿𝗼𝘀𝘀-𝗠𝗼𝗱𝗮𝗹 𝗦𝗮𝗺𝗽𝗹𝗲 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝗼𝗻 — aligning proteins, DNA/RNA, cell data, and images to accelerate multimodal model training. “𝗠𝗶𝗹𝘃𝘂𝘀 𝗲𝗻𝗮𝗯𝗹𝗲𝘀 𝘂𝘀 𝘁𝗼 𝗽𝘂𝘀𝗵 𝘁𝗵𝗲 𝗯𝗼𝘂𝗻𝗱𝗮𝗿𝗶𝗲𝘀 𝗼𝗳 𝘄𝗵𝗮𝘁’𝘀 𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲 𝗶𝗻 𝗯𝗶𝗼𝗹𝗼𝗴𝗶𝗰𝗮𝗹 𝗔𝗜 — 𝘀𝗰𝗮𝗹𝗶𝗻𝗴 𝘀𝗰𝗶𝗲𝗻𝗰𝗲 𝘄𝗶𝘁𝗵 𝘀𝗽𝗲𝗲𝗱 𝗮𝗻𝗱 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆.” — xiaoming zhang at 𝗕𝗶𝗼𝗠𝗮𝗽 🌍 We’re honored to stand with BioMap in pushing the boundaries of science — proving how AI can accelerate discoveries that may change lives. 🔗 Full story: https://lnkd.in/g9AH3uNZ Try Zilliz Cloud for free: https://lnkd.in/g8HMGZXe #VectorDatabase #AI #LifeSciences #BuiltwithMilvus

Vector search should be lightning-fast. But how fast is "fast enough"? 📊 𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 𝘆𝗼𝘂 𝘀𝗵𝗼𝘂𝗹𝗱 𝗸𝗻𝗼𝘄: • < 𝟯𝟬𝗺𝘀 → Target • > 𝟭𝟬𝟬𝗺𝘀 → Investigate • 𝟭 𝘀𝗲𝗰𝗼𝗻𝗱 → Critical 🛠️ 𝗧𝘄𝗼 𝗧𝗼𝗼𝗹𝘀 𝘁𝗼 𝗗𝗶𝗮𝗴𝗻𝗼𝘀𝗲: 1. 𝗠𝗶𝗹𝘃𝘂𝘀 𝗠𝗲𝘁𝗿𝗶𝗰𝘀 (𝗚𝗿𝗮𝗳𝗮𝗻𝗮): Real-time dashboards showing where time is spent—queue, query, or reduce phase. 2. 𝗠𝗶𝗹𝘃𝘂𝘀 𝗟𝗼𝗴𝘀: Automatically logs requests that exceed 1 second, including complete trace information. Know more about Milvus Metrics: https://lnkd.in/gU8XBUmQ Know more about Milvus Logs: https://lnkd.in/gAFyqv5H Next: we'll break down the four most common causes of slow searches and how to fix them. 𝗦𝘁𝗮𝘆 𝘁𝘂𝗻𝗲𝗱! ——— 👉 Follow Milvus, created by Zilliz, for everything related to unstructured data!

𝗜𝗩𝗙 (𝗜𝗻𝘃𝗲𝗿𝘁𝗲𝗱 𝗙𝗶𝗹𝗲 𝗜𝗻𝗱𝗲𝘅) is like organizing a massive library by topic before searching - instead of checking every book, you only look in relevant sections. 𝗧𝗵𝗲 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗣𝗿𝗼𝗰𝗲𝘀𝘀: 1️⃣ 𝗸-𝗺𝗲𝗮𝗻𝘀 𝗖𝗹𝘂𝘀𝘁𝗲𝗿𝗶𝗻𝗴: Groups similar vectors into clusters with centroids 2️⃣ 𝗕𝘂𝗶𝗹𝗱 𝗜𝗻𝘃𝗲𝗿𝘁𝗲𝗱 𝗟𝗶𝘀𝘁: Creates inverted lists mapping each cluster to its member vectors 3️⃣ 𝗤𝘂𝗮𝗻𝘁𝗶𝘇𝗮𝘁𝗶𝗼𝗻: Optionally compresses vectors (PQ/SQ) for 4-64x memory savings 4️⃣ 𝗙𝗶𝗻𝗮𝗹𝗶𝘇𝗲 𝘁𝗵𝗲 𝗜𝗻𝗱𝗲𝘅: Combines centroid table + inverted lists + encoder for the final index 𝗜𝗩𝗙 𝘃𝘀 𝗛𝗡𝗦𝗪: 𝗤𝘂𝗶𝗰𝗸 𝗚𝘂𝗶𝗱𝗲 𝗖𝗵𝗼𝗼𝘀𝗲 𝗜𝗩𝗙: Fast building, memory-efficient, heavy filtering 𝗖𝗵𝗼𝗼𝘀𝗲 𝗛𝗡𝗦𝗪: Ultra-low latency, minimal filtering 💡IVF and HNSW both excel in high-QPS, latency-sensitive applications where performance investment is justified. 𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆 Wrong index choice = 10x higher costs. Understanding your data patterns and query needs is crucial for optimal performance. Learn the details: https://lnkd.in/gnz6TD4S