Zilliz CTO: Why Improving Agent Retrieval Still Requires Export/Reimport of Data
No_Engineer_1224 · reddit · 2026-11-11
James Luan, CTO of Zilliz (behind the open-source Milvus vector database), explains the architectural pain in agent retrieval systems:
- CS/CD loop: He proposes Continuous Serving + Continuous Discovery — Serving handles ingestion and retrieval for online agents, memory, and RAG; Discovery does clustering, dedup, re-embedding, offline eval, and agent trajectory analysis.
- The handoff problem: Fragmented stacks force teams to export vectors to batch environments, rebuild non-reusable indexes, maintain sync, and reimport results, with no shared atomic publication boundary — so discovery work gets postponed.
- Direction: Vector Lakebase extends Milvus toward lake-native continuous discovery, aiming for a repeatable feedback path from offline improvement back into production serving.
More from coding & agent
- AiOS talk recording: helping parallel agents work better with Xcode tooling — Scobleizer · 2026-11-11
- DSPy sandboxes can now import DSPy itself, with warm interpreter pooling — lateinteraction · 2026-11-11
- Fable 5.2 vibe-coding test: two prompts yield a photoreal FPS and a 3D Battle Breakout — vista8 · 2026-11-11
- Nine Engineers, 1,211 PRs in a Day: Lessons From Building Copilot Cowork — usamawahabkhan · 2026-11-11
- jieba-mcp: Chinese Word Segmentation for LLMs via MCP — modelcontextprotocol · 2026-11-11
- Jira MCP Pack Listed: Bring Project Management into AI Clients — modelcontextprotocol · 2026-11-11