Shanghai Meetup Highlights: Xinye's Cost Cuts, Suanzhi Future's Hybrid Search, and AgentSeek's One-Stop Agent Engineering
On May 30, the offline technical salon “Building Highly Reliable, Low-Cost Enterprise-Grade Agent Infra,” co-hosted by OceanBase and the LangChain Community, wrapped up at Zhangjiang Gate of Science in Shanghai. The event brought together frontline practitioners in databases, large models, and agent development, along with enterprise tech leads and open-source community developers, for in-depth exchange around core topics such as building an agent-native data foundation and breaking down real-world industry cases.
Foundational Infrastructure: A Unified Multi-modal Data Foundation
Feng Zhongyan, head of OceanBase open source, kicked off the first keynote, pointing directly at the many problems that arise when most enterprises rely on the file system to carry agent memory: it’s easy to get started early on, but as data volume grows, you quickly run into file redundancy explosions, sluggish retrieval, and an inability to govern data uniformly. He predicted that agent context and memory management will ultimately move toward a unified data foundation that is multi-modal, governable, retrievable, and self-evolving.
Built on the PowerMem memory engine and the seekdb unified data foundation, OceanBase has demonstrated strong results in real-world benchmarks. In tests on LoCoMo and AppWorld: QA accuracy up 65.9%, P95 latency down 91.6%, and token waste reduced 96.5%; the completion pass rate for complex tasks rose from 24% to 39%, a 62.5% increase, while also achieving a 32% reduction in token cost and a 34.7% reduction in task execution steps.
Major New Product Debut: AgentSeek Enterprise-Grade Agent Engineering Platform
LangChain & OceanBase Ambassador Zhang Haili officially unveiled the AgentSeek enterprise-grade agent engineering solution. He made it clear: AgentSeek is not a new development framework, but an agent engineering toolkit for the open-source community and enterprise users, whose core mission is to fill the production-grade gaps in the LangChain ecosystem around serving, deployment, context governance, and the data foundation.
Zhang Haili gave a deep dive into the core philosophy of agent engineering: agent development isn’t a one-time write, but a continuous build → ship → observe → refine → repeat iteration loop. AgentSeek features a five-layer full-stack architecture: the data foundation layer (OceanBase/seekdb), the context semantic layer (ContextSeek), the runtime layer (agentseek-api), the IM gateway layer, and the application layer. All four projects are fully open-sourced and live on GitHub.
Financial Benchmark Practice: Xinye Technology (PPDai) Fully Upgrades to OceanBase
Xia Ping, head of databases at Xinye Technology, shared practical experience on upgrading the database architecture for financial-grade core services. The choice of OceanBase came down to three core advantages: the LSM-Tree high-compression engine for extreme cost reduction, the native distributed architecture that fully retires sharding, and financial-grade Paxos high availability that meets compliance must-haves.
The implementation achieved three key breakthroughs: historical cold-data storage cost dropped about 70%, from 89TB down to 29TB; a same-city active-active and standby-tenant disaster-recovery system was built, achieving RPO=0 and RTO<8 seconds; multi-tenant resource utilization improved 40%, and the new-business deployment cycle was shortened 90%.
AI Hybrid Search Upgrade: Suanzhi Future’s Unified Search Architecture
Chen Song, a database expert at Suanzhi Future, used the synthesis of training corpus for a life-sciences large model as a case study to break down in depth how to build a unified scalar + full-text + regex + vector retrieval foundation on OceanBase. The raw corpus exceeded 20TB, with 3 billion rows of data and 14,000+ files.
The team used OceanBase to transform the massive unstructured JSONL corpus into 9 structured business tables, forming a three-tier retrieval system (L1 exact query, L2 fuzzy/full-text search, L3 vector semantic retrieval). 20TB of files went from scattered, unorganized data to a governable, indexable, and traceable data asset that AI can call directly.
Live Demo: Deploy an Agent with One Command
Shen Honglei, solution director at Jiechuang Intelligence, and Zhang Haili gave two hands-on live demos for the personal digital assistant and the data-analysis deep agent scenarios, respectively.
Shen Honglei completed the environment setup with a single command — no environment configuration, no dependency wrangling. AgentSeek supports smooth scaling from personal → team → enterprise, with a unified database storage layer, so the same architecture upgrades seamlessly from a local demo to an enterprise-grade OceanBase cluster.
The event was packed with substance throughout. Going forward, OceanBase will partner with LangChain to visit more cities, continuing to focus on hands-on topics such as the fusion of agents and data, and context engineering.
👉 Event recap video: https://open.oceanbase.com/activities/4923992