Zilliz

𝗔 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗵𝗮𝘀 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝗼𝗻𝗲 𝗷𝗼𝗯. Teams need low-latency retrieval for live applications. But they also need to explore datasets, run bulk operations, and find near-duplicate content at scale. That requires more than fast top-K search. It requires a system that can support different workloads without forcing teams to assemble a new data stack for each one. This is the direction behind Zilliz Cloud: 🔎 𝗢𝗻𝗹𝗶𝗻𝗲 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹: dense, sparse, and hybrid search, with metadata filtering and reranking to improve relevance. 🧭 𝗗𝗮𝘁𝗮 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝘆: range, grouping, and filtered search help teams examine relationships in large vector datasets—not only return the nearest few results. 🧹 𝗗𝗲𝗱𝘂𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻: MinHash functions enable near-duplicate detection across large text collections. Our CTO, Xiaofan(James) Luan, discussed the broader shift in an earlier Innovator Coffee conversation: vector databases are becoming a production data layer for AI, not simply a similarity-search tool. That is the future we are building toward with Zilliz Cloud. ☁️ Read the full story 👇 https://lnkd.in/gTn8XWry

During our recent webinar, we received a great question: “Do you expect more vector databases to move toward a disaggregated storage and compute architecture?” Simon Hearne’s answer: yes—but vector lakebase and real-time vector serving solve different problems. Using autonomous driving as an example, he explains why teams may need both: one for analyzing fleet-wide data and retraining models, and the other for low-latency operational queries. Watch the video clip for Simon’s full explanation. 👇 Full webniar 👉🏻: https://lnkd.in/gii3Mrbw

"Can you keep our data in Japan?" "Do you have anything in Canada?" "Is Frankfurt available?" These come up in almost every enterprise conversation. So we put the answer in one place. 🌍 Zilliz Cloud is now available in 𝟮𝟰 𝗰𝗹𝗼𝘂𝗱-𝗿𝗲𝗴𝗶𝗼𝗻 𝗰𝗼𝗺𝗯𝗶𝗻𝗮𝘁𝗶𝗼𝗻𝘀 across AWS, Google Cloud, and Azure. Virginia, Ohio, Oregon, Iowa, Canada Central, Frankfurt, Ireland, London, Tokyo, Seoul, Singapore, Hong Kong SAR, Pune, Sydney. Virginia and Frankfurt are on all three clouds. Not seeing the one you need?  Request it on the same page — that's how we decide what comes next. 👉🏻 https://lnkd.in/gn5BXDdT #VectorDatabase #AIInfrastructure

Zilliz August recap: we focused on four areas: a stronger data engine, clearer operations, flexible deployment, and real-world proof. 𝟭.𝗦𝘁𝗿𝗼𝗻𝗴𝗲𝗿 𝗱𝗮𝘁𝗮 𝗲𝗻𝗴𝗶𝗻𝗲 ⚙️ Need more than vector similarity?  Milvus 3.0 added Snapshots, schema evolution, aggregation, ORDER BY, and enhanced FTS, reducing recovery, migration, and application-side processing. 👉 https://lnkd.in/gyd2nRcq 𝟮.𝗖𝗹𝗲𝗮𝗿𝗲𝗿 𝗼𝗽𝗲𝗿𝗮𝘁𝗶𝗼𝗻𝘀 🔍 Need workload visibility?  Collection metrics show QPS, latency, and capacity by collection, helping isolate production issues. 𝟯.𝗙𝗹𝗲𝘅𝗶𝗯𝗹𝗲 𝗱𝗲𝗽𝗹𝗼𝘆𝗺𝗲𝗻𝘁 🌍 Need locality or infrastructure control?  London and Tokyo regions, BYOC-I on Google Cloud, and on-demand BYOC expand deployment choice. 𝟰.𝗣𝗿𝗼𝘃𝗲𝗻 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲 🧪 Need evidence and implementation guidance? -> Creative search at scale — Need prompts and assets searchable together? OpenArt cut P99 from 25s to ~300ms and removed its custom fusion layer. -> Persistent agent memory — Need context across sessions and tools? MemSearch × DeepSeek Harness keeps relevant project context available. -> Enhanced full-text search — Need proof of stronger keyword relevance? The webinar compared Milvus with Elasticsearch across lexical, semantic, and hybrid search. 👉 https://lnkd.in/gii3Mrbw -> Architecture guidance — Need help with your own stack? Office Hours offered 1:1 guidance on schema, performance, scaling, and troubleshooting. 👉 https://lnkd.in/gAxPnxpq Which part creates the most friction? #Milvus #ZillizCloud

💾 How much memory does hybrid search really need? 𝗜𝗻 𝘁𝗵𝗶𝘀 𝗰𝗹𝗶𝗽 𝗳𝗿𝗼𝗺 𝗼𝘂𝗿 𝗠𝗶𝗹𝘃𝘂𝘀 𝟯.𝟬 𝘄𝗲𝗯𝗶𝗻𝗮𝗿, @𝗦𝗶𝗺𝗼𝗻𝗲 𝗛𝗲𝗮𝗿𝗻𝗲 𝘄𝗮𝗹𝗸𝘀 𝘁𝗵𝗿𝗼𝘂𝗴𝗵 𝗮 𝗹𝗶𝗸𝗲-𝗳𝗼𝗿-𝗹𝗶𝗸𝗲 𝗹𝗼𝗰𝗮𝗹 𝗗𝗼𝗰𝗸𝗲𝗿 𝘁𝗲𝘀𝘁 𝗰𝗼𝗺𝗽𝗮𝗿𝗶𝗻𝗴 𝗠𝗶𝗹𝘃𝘂𝘀 𝟯.𝟬 𝘄𝗶𝘁𝗵 𝗘𝗹𝗮𝘀𝘁𝗶𝗰𝘀𝗲𝗮𝗿𝗰𝗵. Both were capped at 8 GB RAM and 4 CPUs, then loaded with the same 100K Amazon review documents, vectors, and BM25 setup. The result: Milvus 3.0 used about 0.5 GB of memory, including its subprocesses, while Elasticsearch used about 5.2 GB. It’s a laptop-scale test, not a universal production benchmark—but the gap is worth a closer look. ▶️ Watch the full Milvus 3.0 webinar here: https://lnkd.in/gii3Mrbw

📢 𝗡𝗲𝘄 𝘄𝗲𝗯𝗶𝗻𝗮𝗿: 𝗠𝗶𝗴𝗿𝗮𝘁𝗶𝗻𝗴 𝗳𝗿𝗼𝗺 𝗘𝗹𝗮𝘀𝘁𝗶𝗰𝘀𝗲𝗮𝗿𝗰𝗵/𝗢𝗽𝗲𝗻𝗦𝗲𝗮𝗿𝗰𝗵 𝘁𝗼 𝗠𝗶𝗹𝘃𝘂𝘀 🗓️ Sep 2, 2026 · 11 AM PT / 2 PM ET · with Simon Hearne, Solutions Architect Zilliz After a packed first session comparing Elasticsearch and Milvus 3.0 on full-text search, Simon moves from comparison to 𝗺𝗶𝗴𝗿𝗮𝘁𝗶𝗼𝗻 — hands-on, with a real end-to-end walkthrough. 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂'𝗹𝗹 𝗴𝗲𝘁: → The migration path from ES/Opensearch to Milvus 3.0 — key stages and decisions, start to finish → A live walkthrough with both open-source VTS and its managed vector migration tools  → Elastic Cloud (ESS) vs. Zilliz Cloud (managed Milvus) — licensing & commercial tradeoffs → Deployment choices — open-source Milvus, Zilliz Cloud, and Zilliz BYOC → Customer migration stories — why teams moved, and what they learned → A ~20-min live AMA with Simon Running Elasticsearch or OpenSearch and eyeing a move? This is the practical one. 👉 Save your spot: https://lnkd.in/g7G4Jt2M Thinking about moving off ES/OpenSearch? What's been the blocker? Leave your comments! ——— 👉 Follow Zilliz for vector database and vector lakebase updates built for production AI. #vectordatabase #vectorlakebase #elasticsearch

For agent workloads, we built Woodpecker to make the WAL cheaper to run and easier to scale. Agents keep writing. Memory updates, new context, fresh embeddings, background re-embedding. At scale, the write path starts to matter almost as much as storage and query performance. Traditional node-bound WALs make writes compete with indexing and queries. Kafka and Pulsar decouple the log, but add brokers, replicated storage, and more infrastructure to operate. Woodpecker takes a more cloud-native path. Embedded mode batches logs and writes them directly to object storage, which fits bulk ingestion, periodic syncs, and historical backfills. Service mode is built for real-time ingestion. It keeps the WAL separate from database compute, while shortening the data path with client-driven replication, topology-aware reads, and earlier offloading of rolled segments to object storage. The cost difference can be substantial. At a sustained 20 MB/s write rate with 72 hours of log retention, our AWS pricing estimate puts three-replica gp3 storage at about $1,800/month. Moving rolled segments to S3 brings the corresponding storage cost to about $115/month.

🎬 𝗧𝘆𝗽𝗲 𝗮 𝗰𝗵𝗮𝗿𝗮𝗰𝘁𝗲𝗿, 𝗺𝗼𝗼𝗱, 𝗼𝗿 𝘀𝗰𝗲𝗻𝗲. 𝗙𝗶𝗻𝗱 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗺𝗼𝗺𝗲𝗻𝘁 𝗳𝗿𝗼𝗺 𝗺𝗼𝗻𝘁𝗵𝘀 𝗼𝗳 𝗔𝗜 𝗰𝗿𝗲𝗮𝘁𝗶𝗼𝗻𝘀. For OpenArt AI’s 8M+ creators, generation history is more than an archive. Every past image and video can become material for the next character, scene, or story. But making that history searchable became increasingly difficult. Elasticsearch queries fanned out across a growing number of indices, pulled data from frozen storage, and reached 25 seconds at P99. Searching both the creator’s prompt and the generated asset also required a hand-built dual-kNN fusion layer. OpenArt moved the workload to Zilliz Cloud. Each creation is now represented by two signals: a prompt vector for what the creator requested and an asset vector for what the model produced. Zilliz Cloud searches both together in one weighted request. 🚀 𝗞𝗲𝘆 𝗥𝗲𝘀𝘂𝗹𝘁𝘀 ⚡ 𝗥𝗼𝘂𝗴𝗵𝗹𝘆 𝟯𝟬𝟬 𝗺𝘀 𝗮𝘁 𝗣𝟵𝟵, down from 25 seconds 💰 𝗥𝗼𝘂𝗴𝗵𝗹𝘆 𝟴𝟱% 𝗹𝗼𝘄𝗲𝗿 𝗰𝗼𝗺𝗽𝘂𝘁𝗲 𝗰𝗼𝘀𝘁 📦 𝗥𝗼𝘂𝗴𝗵𝗹𝘆 𝟰𝟱𝟲 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝘃𝗲𝗰𝘁𝗼𝗿𝘀 migrated without re-embedding 🧩 𝗡𝗼 𝗺𝗼𝗿𝗲 𝗰𝘂𝘀𝘁𝗼𝗺 𝗳𝘂𝘀𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿 to maintain What was once a growing infrastructure project is now a fast, meaning-based search experience. A creator’s entire library can function as one creative memory. 🔗 Full story: https://lnkd.in/gwKDYKzC 👉 Try Zilliz Cloud Free: https://zilliz.com/cloud ——— 👉 Follow Zilliz for vector database and vector lakebase updates built for production AI. #ScaleWithZilliz #VectorDatabase #AICreators

Switching to a new embedding model can look like a simple API change. In production, it can become a full data migration. Moving from single-vector to multi-vector retrieval can introduce an even larger architectural change. That means the real question is no longer simply, “Which model has the best benchmark score?” It is whether the model can: • Improve retrieval of your real queries • Fix failure cases that matter • Deliver enough value to justify the operational cost of switching Our latest guide explains how to evaluate embedding models for the second half of 2026:

📣📣 𝗭𝗶𝗹𝗹𝗶𝘇 𝗖𝗹𝗼𝘂𝗱 𝗶𝘀 𝗻𝗼𝘄 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲 𝗶𝗻 𝘁𝗵𝗲 𝗔𝗪𝗦 𝗧𝗼𝗸𝘆𝗼 𝗿𝗲𝗴𝗶𝗼𝗻 (𝗮𝗽-𝗻𝗼𝗿𝘁𝗵𝗲𝗮𝘀𝘁-𝟭)! As AI applications reach more users across Japan, organizations need infrastructure that can deliver low latency, support local data residency, and control costs. What this means to you: 🌍 𝗟𝗼𝘄𝗲𝗿 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 - Run vector search closer to users in Japan for faster, more consistent responses. 🔒 𝗝𝗮𝗽𝗮𝗻 𝗱𝗮𝘁𝗮 𝗿𝗲𝘀𝗶𝗱𝗲𝗻𝗰𝘆 - Keep vector data in Japan to support residency and compliance requirements. 💰 𝗟𝗼𝘄𝗲𝗿 𝗱𝗮𝘁𝗮 𝘁𝗿𝗮𝗻𝘀𝗳𝗲𝗿 𝗰𝗼𝘀𝘁𝘀 - Co-locate your vector database with Japan-based applications to reduce cross-region traffic and related transfer costs. ⚡ 𝗦𝘁𝗿𝗼𝗻𝗴𝗲𝗿 𝗔𝗜 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 - Support more responsive RAG, search, chat, and recommendation systems. Zilliz Cloud now spans 𝟮𝟯 regions across AWS, Google Cloud, and Microsoft Azure, giving teams more flexibility to deploy closer to their users, data, and applications. 👉 See all available regions: https://lnkd.in/eABfbWHQ