Milvus Meetup - San Francisco
Running a more powerful retrieval engine in production
Milvus has expanded beyond vector search into a more powerful retrieval engine—combining vector, full-text, and structured search over data stored in open lake formats.
Milvus 3.0 brings more of the retrieval workflow into one engine while allowing data to remain in formats such as Parquet, Lance, Iceberg, and Vortex—without maintaining another serving copy.
Join Milvus maintainers, AWS architects, and engineers working on retrieval at scale for a practical look at the new architecture, its tradeoffs, and what it changes in production.
What we’ll cover
Vector, full-text, and structured search in one retrieval engine
Searching lake-resident data without another serving copy
When to use External Collections versus native Milvus collections
Using snapshots and batch workflows for evaluation and re-embedding
Agenda
17:00 — Check-in. Food and drinks.
17:30 — Opening: AWS welcome + Milvus community intro
17:40
Talk: Milvus 3.0: A More Powerful Retrieval Engine
Speaker: Jiang Chen, Head of Developer Relations and Solutions, Milvus/Zilliz
Outline:
Engineers maintaining a data lake and retrieval system faces a challenge: either copy lake-resident embeddings into a separate serving system, or search the lake directly and accept brute-force scans and impractical latency.
Milvus 3.0 introduces a third path.
With lake-native External Collections, Milvus can build index thus efficiently serve searches on vector, full-text, JSON, and scalar data over lake files on object storage. Whether it’s Parquet, Lance, Iceberg, or Vortex format, the source data stays where it is, without another serving copy or an ETL pipeline that must constantly be kept in sync.
But eliminating a copy is only part of the story. Milvus 3.0 also moves more of the retrieval pipeline into the engine: sorting, aggregation, faceted search, nested multi-vector data, late-interaction retrieval, and a redesigned sparse index.
This talk will share the architectural change that makes this possible, as well as new features in full-text search and more flexible data modeling.
18:10 — [TBD-Talk title] [TBD-Speaker name], [TBD-Company]
18:40 — Break. Networking.
19:00 — [TBD-Talk title] [TBD-Speaker name], [TBD-Company]
19:30 — [TBD-Talk title] [TBD-Speaker name], Amazon Web Services (AWS)
20:00 — Panel and open Q&A with all speakers
20:20 — Networking
21:00 — Close
Who this is for
Search, ML infrastructure, and data platform engineers
AI architects, startup CTOs, and founding engineers
Teams building or evaluating production retrieval systems
A technical evening focused on modern architectures, production operations, and real-world tradeoffs. Bring hard questions, please.
Before you come
Registration closes 14 October and late registrations cannot be admitted — reception works from a list submitted three business days ahead, and anyone not on it will not get into the building. Please bring photo ID. Check-in is two stages. Detailed entrance directions go out by email once your registration is confirmed.