登录

milvus

milvus-ioApache-2.0活跃维护

高性能云原生向量数据库，专为大规模向量近似最近邻检索设计，适合搭建以图搜图、语义搜索和 RAG 等应用。

Milvus is a high-performance, cloud-native vector database built for scalable vector ANN search

v3.0.5发布于 2026-10-30 10:43 UTC46,632star +20 / 7 天4,327fork同步于 2026-11-10 23:20 UTC
向量数据库近似最近邻检索云原生分布式RAG
对比 GitHub 官网annscloud-nativediskanndistributedembedding-database

怎么上手

由 README 提炼，执行前请确认
原文 
pip install -U pymilvus
pip install pymilvus[milvus-lite]
硬件要求按 README 说明

README 未说明具体显存或 GPU 型号要求；提到对 CPU/GPU 做硬件加速，并支持 Standalone 单机部署与 Milvus Lite 轻量本地模式。

支持 Apple Silicon
下载：源码 zip·GitHub Releases（安装包）本站不镜像安装包

项目说明

由 README 自动翻译提炼
没人工点评？写下你的经验

高性能云原生向量数据库，为十亿级向量的可扩展 ANN 检索而设计

高性能云原生向量数据库，专为大规模向量近似最近邻检索设计，适合搭建以图搜图、语义搜索和 RAG 等应用。

核心能力

  • 计算与存储分离的分布式架构，可横向扩展
  • 支持 HNSW、IVF、FLAT、SCANN、DiskANN 等多种索引
  • 支持元数据过滤与混合搜索
  • 提供 Milvus Lite，用 pip 装完即可本地起库
  • 对 CPU/GPU 做硬件加速优化向量检索

容易踩的坑

Milvus Lite 需额外安装 pymilvus[milvus-lite] 扩展包，只装 pymilvus 无法本地起库；从源码编译要求 CMake >= 3.26.4 且 < 4、Python > 3.8 且 <= 3.11，版本区间卡得较死；连接自建服务或 Zilliz Cloud 需自行准备 endpoint 与 token。

内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README

最近发版

5 个版本v3.0.5
全部 release
  1. v3.0.5milvus-3.0.5最新正式版2026-10-30 · 2026-10-30 10:43 UTC

    v3.0.5

    Release date: October 30, 2026

    Milvus VersionPython SDK VersionNode.js SDK VersionJava SDK VersionGo SDK Version
    3.0.53.0.23.0.63.0.123.0.0

    We are excited to announce the release of Milvus v3.0.5! This release adds resource-group-aware load status in QueryCoord, reduces memory use in several paths. It also fixes stability issues around coordinator restarts, rolling upgrades and transient etcd errors.

    Improvements

    • Added per-resource-group load percentage and shard-leader readiness in QueryCoord, and tagged each shard leader with its replica's resource group, so queries can be routed within a single resource group (#52500, #52711)
    • Reduced copying in unscored Tantivy term queries by borrowing decoded postings blocks directly (#54469)
    • Upgraded Knowhere to v3.0.17 and moved nullable vector offset mapping behind Knowhere APIs (#54473)
    • Reduced memory allocations when encoding batched RESTful query and search responses by reusing per-request encoding buffers (#54540)
    • Improved text analyzer performance by reusing compiled regexes and per-filter search caches in analyzer filters (#54570)
    • Added an opt-in two-tier mix compaction planner (dataCoord.compaction.twoTierCompaction, default off) to reduce compaction write amplification (#54589)
    • Reduced timestamp index memory usage when loading small sealed segments (#54659)
    • Added the missing status="fail" label to the internal_storage_op_count metric for failed get operations; the old unlabeled series stops and the new series starts from zero (#54637)
    • Allowed hyphens in resource group names after the first character, matching the naming rules for partitions (#54679)
    • Reduced memory allocations in compaction merge sort by reusing output columns across records (#54718)

    Bug fixes

    • Fixed an issue where collections with struct array fields could force-load raw embedding-list vectors on first search or after cache eviction, increasing memory usage and query latency (#54167)
    • Fixed an issue where upserts on CMEK-enabled collections wrote the delete part (primary keys and timestamps) to the WAL without encryption (#54478)
    • Fixed an issue where ST_DWITHIN filters returned fewer rows on segments with an RTREE index than on unindexed segments; empty geometries no longer match ST_DWITHIN, consistent with PostGIS (#54611)
    • Fixed an issue where writes to Storage V3 segments could stall after a coordinator restart because the next flush used a column layout that conflicted with the committed manifest (#54652)
    • Fixed an issue where QueryNode could panic while replaying encrypted WAL messages from a dropped database that shared a physical channel with other collections (#54640)
    • Fixed an issue where partial upserts failed on collections with nullable struct array fields when the field was omitted and the batch contained new primary keys (#54619)
    • Fixed an issue where index state was incorrectly reported
  2. v3.0.4milvus-3.0.42026-10-26 · 2026-10-26 06:53 UTC
    查看更新说明

    v3.0.4

    Release date: October 26, 2026

    Milvus VersionPython SDK VersionNode.js SDK VersionJava SDK VersionGo SDK Version
    3.0.43.0.23.0.63.0.103.0.0

    We are excited to announce the release of Milvus v3.0.4! This release speeds up bulk import, compaction, and filtered and grouped search, adds lazy column-group loading for Storage V3, and fixes hangs in import, collection loading, and partial-update writes, along with several access control issues and CVEs.

    Improvements

    • Improved sort compaction read throughput by reading input chunks in parallel and prefetching byte ranges eagerly (#53437)
    • Added support for precomputed sparse vectors in external collections backed by Parquet, Lance, and Vortex files (#53894)
    • Refactored Proxy DML tasks into a dedicated package to separate task responsibilities (#54064)
    • Ensured all output fields of a search on sealed segments are filled from the same segment snapshot as the search itself (#54074)
    • Added the queryNode.segcore.tieredStorage.lazyColumnGroupEnabled option (disabled by default) to defer loading Storage V3 column groups until first access, reducing segment load time and memory usage (#54169)
    • Added autoIndex.params.sparse.ip.build and autoIndex.params.sparse.bm25.build to configure sparse AutoIndex build parameters separately for IP and BM25 metrics, falling back to autoIndex.params.sparse.build (#54212)
    • Reduced health check latency on clusters with large segment metadata by no longer collecting DataCoord self metrics during CheckHealth (#54214)
    • Improved DataCoord import scheduling performance for clusters with many import jobs by indexing import tasks by job ID (#54202)
    • Reduced object storage list requests during garbage collection by scanning obsolete JSON stats prefixes once per GC cycle instead of once per segment (#54207)
    • Improved ARRAY_LENGTH filter performance on sealed segments by reading lengths from existing array chunk offsets, raising QPS from 102 to 270 on a customer filter workload (#54321)
    • Upgraded grpc-go to v1.84.0 so Proxy releases raw request buffers right after decoding, saving up to the full request size (61.5 MB for an nq=10000, 1536-dim search) of heap per in-flight large search (#54271)
    • Added a hidden streaming.partialUpdate.enableCAS option (enabled by default) to turn off compare-and-swap checks for partial updates on StreamingNode, trading protection against concurrent stale updates for lower write overhead (#54249)
    • Removed a spurious "status is nil" warning that Proxy logged on every successful query (#54247)
    • Reduced memory allocations when decoding RESTful V2 insert and upsert requests, and stopped logging full request payloads on conversion failures (#54403)
    • Reworked bulk import to assign exact per-file ID ranges after preimport, so import requests return immediately without sizing files up front, and fixed log ID exhaustion during import write retries (#54416)
    • Added startup configuration options to override the interim index algorithm for IP and BM25 sparse vecto
  3. v2.6.26milvus-2.6.262026-10-26 · 2026-10-26 06:53 UTC
    查看更新说明

    v2.6.26

    Release date: October 26, 2026

    Milvus VersionPython SDK VersionNode.js SDK VersionJava SDK VersionGo SDK Version
    2.6.262.6.172.6.172.6.272.6.5

    We are excited to announce the release of Milvus v2.6.26! This release speeds up RPC message decoding, cuts Proxy memory usage for large search and query requests, and fixes QueryNode crashes in nullable and NGRAM filtering, hangs during collection loading and import jobs, and timestamps that could go backward after coordinator startup.

    Improvements

    • Sped up decoding of search, query, insert, and upsert RPC messages with a faster protobuf decoder, enabled by default and switchable via common.enableFastPB (#53994, #54002)
    • Pinned the text analyzer stemmer independently of Tantivy so that Tantivy upgrades do not change stemming results or cause missed matches on existing text indexes and BM25 data (#54029, #54098)
    • Reduced the delay between an async segcore operation completing and its caller receiving the result under heavy load, so completed results no longer wait behind queued work while holding memory (#54128)
    • Reduced Proxy memory usage for large search and query requests by sharing request payloads across shards instead of deep-copying them per shard; live heap dropped from 8.7 GB to 5.85 GB in a test with 50 concurrent nq=10000 searches (#54224)
    • Merged same-column array_contains conditions in OR/AND chains into array_contains_any/array_contains_all during plan compilation, reducing compilation overhead for filters with thousands of conditions (#54435)
    • Improved null-check filtering on JSON path indexes in sealed segments by caching the validity bitmap instead of rescanning the index for every query (#54422)
    • Added the queryNode.segcore.interimIndex.sparseIPIndexAlgo and sparseBM25IndexAlgo settings to choose the interim index algorithm (TAAT_NAIVE, DAAT_WAND, or DAAT_MAXSCORE) for sparse IP and BM25 search on growing and unindexed sealed segments (#54443)
    • Reduced repeated Parquet footer reads and parsing when loading StorageV2 segments in multiple batches, and picked up a milvus-storage fix for the endpoint URL format on Azure sovereign clouds (#54488)

    Bug fixes

    • Fixed an issue where grouping search with strict group size could return groups with fewer than group_size results even when enough matching entities existed; added queryNode.groupBy.strictGroupStrategy and related settings to tune this behavior (#53451)
    • Fixed an issue where filtering on a nullable field while inserts or upserts were running could crash the QueryNode or misread null values in growing segments (#53870)
    • Fixed an issue where filters that combine AND/OR with out-of-range integer comparisons on nullable integer fields could crash the QueryNode or return incorrect results (#53897)
    • Fixed an issue where LIKE queries on fields with an NGRAM index could crash the QueryNode or return non-matching rows when the pattern contained NUL characters (#53710)
    • Fixed an issue where a failed read of the persisted TSO window during coordinator startup could cause newly allocated time
  4. v3.0.3milvus-3.0.32026-10-14 · 2026-10-14 07:58 UTC
    查看更新说明

    v3.0.3

    Release date: October 14, 2026

    Milvus VersionPython SDK VersionNode.js SDK VersionJava SDK VersionGo SDK Version
    3.0.33.0.23.0.63.0.103.0.0

    We are excited to announce the release of Milvus v3.0.3! This release brings Function Chain to Hybrid Search, element-level partial updates for Array, Struct, and JSON fields, and IS NULL filtering on nullable vector fields, along with security hardening and a broad set of stability fixes; note the breaking changes to legacy V1 REST routes on the metrics port, binlog import privileges, and the default Pulsar producer access mode.

    Features

    • Added Function Chain support to Hybrid Search, letting a top-level L2 chain starting with merge define fusion and reranking and letting each sub-search run its own L0/L1 chain before fusion (#53541)
    • Added support for replacing individual Array elements, Struct array elements or their sub-fields, and JSON paths in partial updates, available in the Go SDK and REST v2 (#53746)
    • Added dynamic field support to function chains in search and hybrid search (#54011)
    • Added support for IS NULL and IS NOT NULL filter expressions on nullable vector fields, with batched validity checks for efficient evaluation (#54020)

    Improvements

    • Added optional root authentication for management endpoints on the metrics port (common.security.adminAuthEnabled, disabled by default) and redacted credentials and other sensitive values from configuration dumps, HTTP responses and logs (#53499)
    • Preserved the primary keys of existing entities in full upsert on AutoID collections, assigning new IDs only to keys that are not found; such upserts now require the collection to be loaded (#53511)
    • Added snappy and s2 gRPC compression algorithms, made the zstd codec configurable, and hardened gRPC decompression against oversized allocations (#53598)
    • Blocked import paths that point into Milvus's internal storage directories and required the new cluster-level ImportBinlog privilege for binlog (backup=true) and L0 imports; this breaks non-admin roles that have only the collection-level Import privilege until they are granted ImportBinlog (#53680)
    • Optimized bitset writes for Tantivy-backed index lookups, making writes of contiguous doc IDs about 7x faster in microbenchmarks (#53700)
    • Reduced redundant payload and metadata copies in the C++ query execution path (#53756)
    • Deferred loading of JSON stats for sealed segments until first use when warmup is disabled, so segments whose JSON stats are never queried no longer pay the load cost (#53888)
    • Added an optional QueryCoord precheck that estimates memory and local disk demand before loading collections or partitions, with per-resource-group demand metrics (#53905)
    • Reduced per-chunk overhead in filter expressions and group-by search on sealed segments by reading data through the request-pinned column snapshot (#53984)
    • Accelerated RPC protobuf decoding with the fastPB decoder for search, query, insert, and upsert paths, enabled by default and controllable via common.enableFastPB (#53995, [#540
  5. v2.6.25milvus-2.6.252026-09-29 · 2026-09-29 06:56 UTC
    查看更新说明

    v2.6.25

    Release date: September 29, 2026

    Milvus VersionPython SDK VersionNode.js SDK VersionJava SDK VersionGo SDK Version
    2.6.252.6.172.6.172.6.252.6.5

    We are excited to announce the release of Milvus v2.6.25! This release tightens security with stricter RBAC checks, a new ImportBinlog privilege, and masking of sensitive configuration values. It also adds snappy and s2 gRPC compression, and fixes stability issues around shard leader movement, scalar index replacement, and Pulsar channel takeover.

    Improvements

    • Added per-resource-group load config convergence reporting to the management API, covering replica distribution, serviceability, query visibility, and residual resources (#53518)
    • Added snappy and s2 gRPC compression, made the zstd codec configurable (grpc.client.compressionAlgorithm, compressionLevel, compressionCRC, compressionConcurrency), and hardened zstd decompression against oversized messages (#53599)
    • Masked sensitive values and omitted undeclared keys in configuration views and diagnostic logs, so credentials and unrelated environment variables are no longer exposed (#53613)
    • Required the new cluster-level ImportBinlog privilege for binlog (backup=true) and L0 imports, and rejected regular imports whose file paths point into Milvus's internal storage directories (#53681)
    • Sped up writing Tantivy index results into bitsets, up to 7x faster per doc ID for contiguous IDs in microbenchmarks (#53701)
    • Hardened RBAC enforcement by disabling the Dummy RPC's query mode, which bypassed authorization and rate limiting, and by requiring Search privilege on the collection for RunAnalyzer requests that specify one (#53812)
    • Reduced redundant DescribeCollection RPCs from Proxy to the coordinator during insert, delete, and upsert by resolving collection channels from the Proxy metadata cache (#53793)

    Bug fixes

    • Fixed a DataNode memory leak where a fully built index or analyze result was never released if its task was canceled during the build (#53349)
    • Fixed an issue where loading segments with FLAT or BIN_FLAT indexes could fail during incremental index loading, especially under restrictive resource limits (#53383)
    • Fixed an issue where index builds on segments containing malformed JSON retried indefinitely; such tasks now fail fast with a permanent error instead of looping across workers (#53429)
    • Fixed an issue where replacing a scalar index on a loaded collection could leave the old index loaded indefinitely and cause repeated reloads; both QueryCoord and QueryNode must be upgraded for the complete fix (#53537)
    • Fixed an issue where a StreamingNode could keep writing to a Pulsar-backed channel after another node took it over; pulsar.producerAccessMode now defaults to exclusive (enforcement requires Pulsar 2.8.0+; set it to shared to keep the previous behavior) (#53540)
    • Fixed an issue where sustained GIS query load could stall inserts into growing segments when the geometry cache (enableGeometryCache) was enabled (#53610)
    • Fixed an issue where Milvus components could hang during shutdown or block on logging when the async lo

用户点评

登录用户的真实使用记录，可按版本和硬件筛选
共 0 条

    还没有人点评。第一个说出真实体验的人最有价值。

    README 原文

    未翻译
    在 GitHub 打开

    What is Milvus?

    🐦 Milvus is a high-performance vector database built for scale. It powers AI applications by efficiently organizing and searching vast amounts of unstructured data, such as text, images, and multi-modal information.

    🧑‍💻 Written in Go and C++, Milvus implements hardware acceleration for CPU/GPU to achieve best-in-class vector search performance. Thanks to its fully-distributed and K8s-native architecture, Milvus can scale horizontally, handle tens of thousands of search queries on billions of vectors, and keep data fresh with real-time streaming updates. Milvus also supports Standalone mode for single machine deployment. Milvus Lite is a lightweight version good for quickstart in python with pip install.

    Want to use Milvus with zero setup? Try out Zilliz Cloud ☁️ for free. Milvus is available as a fully managed service on Zilliz Cloud, with Serverless, Dedicated and BYOC options available.

    For questions about how to use Milvus, join the community on Discord to get help. For reporting problems, file bugs and feature requests in GitHub Issues or ask in Discussions.

    The Milvus open-source project is under LF AI & Data Foundation, distributed with Apache 2.0 License, with Zilliz as its major contributor.

    Quickstart

    $ pip install -U pymilvus

    This installs pymilvus, the Python SDK for Milvus. Use MilvusClient to create a client:

    from pymilvus import MilvusClient

    • You can also try Milvus Lite for quickstart by installing pymilvus[milvus-lite]. To create a local vector database, simply instantiate a client with a local file name for persisting data:

      client = MilvusClient("milvus_demo.db")

    • You can also specify the credentials to connect to your deployed Milvus server or Zilliz Cloud:

      client = MilvusClient(
  uri="<endpoint_of_self_hosted_milvus_or_zilliz_cloud>",
  token="<username_and_password_or_zilliz_cloud_api_key>")

    With the client, you can create collection:

    client.create_collection(
    collection_name="demo_collection",
    dimension=768,  # The vectors we will use in this demo have 768 dimensions
)

    Ingest data:

    res = client.insert(collection_name="demo_collection", data=data)

    Perform vector search:

    query_vectors = embedding_fn.encode_queries(["Who is Alan Turing?", "What is AI?"])
res = client.search(
    collection_name="demo_collection",  # target collection
    data=query_vectors,  # a list of one or more query vectors, supports batch
    limit=2,  # how many results to return (topK)
    output_fields=["vector", "text", "subject"],  # what fields to return
)

    Why Milvus

    Milvus is designed to handle vector search at scale. It stores vectors, which are learned representations of unstructured data, together with other scalar data types such as integers, strings, and JSON objects. Users can conduct efficient vector search with metadata filtering or hybrid search. Here are why developers choose Milvus as the vector database for AI applications:

    High Performance at Scale and High Availability

    • Milvus features a distributed architecture that separates compute and storage. Milvus can horizontally scale and adapt to diverse traffic patterns, achieving optimal performance by independently increasing query nodes for read-heavy workload and data node for write-heavy workload. The stateless microservices on K8s allow quick recovery from failure, ensuring high availability. The support for replicas further enhances fault tolerance and throughput by loading data segments on multiple query nodes. See benchmark for performance comparison.

    Support for Various Vector Index Types and Hardware Acceleration

    • Milvus separates the system and core vector search engine, allowing it to support all major vector index types that are optimized for different scenarios, including HNSW, IVF, FLAT (brute-force), SCANN, and DiskANN, with quantization-based variations and mmap. Milvus optimizes vector search for advanced features such as metadata filtering and range search. Additionally, Milvus implements hardware acceleration to enhance vector search performance and supports GPU indexing, such as NVIDIA's CAGRA.

    Flexible Multi-tenancy and Hot/Cold Storage

    • Milvus supports multi-tenancy through isolation at database, collection, partition, or partition key level. The flexible strategies allow a single cluster to handle hundreds to millions of tenants, also ensures optimized search performance and flexible access control. Milvus enhances cost-effectiveness with hot/cold storage. Frequently accessed hot data can be stored in memory or on SSDs for better performance, while less-accessed cold data is kept on slower, cost-effective storage. This mechanism can significantly reduce costs while maintaining high performance for critical tasks.

    Sparse Vector for Full Text Search and Hybrid Search

    • In addition to semantic search through dense vector, Milvus also natively supports full text search with BM25 as well as learned sparse embeddings such as SPLADE and BGE-M3. Users can store sparse vectors and dense vectors in the same collection, and define functions to rerank results from multiple search requests. See examples of Hybrid Search with semantic search + full text search.

    Data Security and Fine-grain Access Control

    • Milvus ensures data security by implementing mandatory user authentication, TLS encryption, and Role-Based Access Control (RBAC). User authentication ensures that only authorized users with valid credentials can access the database, while TLS encryption secures all communications within the network. Additionally, RBAC allows for fine-grained access control by assigning specific permissions to users based on their roles. These features make Milvus a robust and secure choice for enterprise applications, protecting sensitive data from unauthorized access and potential breaches.

    Milvus is trusted by AI developers to build applications such as text and image search, Retrieval-Augmented Generation (RAG), and recommendation systems. Milvus powers many mission-critical businesses for startups and enterprises.

    Demos and Tutorials

    Here is a selection of demos and tutorials to show how to build various types of AI applications made with Milvus:

    You can explore a comprehensive Tutorials Overview covering topics such as Retrieval-Augmented Generation (RAG), Semantic Search, Hybrid Search, Question Answering, Recommendation Systems, and various quick-start guides. These resources are designed to help you get started quickly and efficiently.

    TutorialUse CaseRelated Milvus Features
    Build RAG with MilvusRAGvector search
    Advanced RAG OptimizationsRAGvector search, full text search
    Full Text Search with MilvusText Searchfull text search
    Hybrid Search with MilvusHybrid Searchhybrid search, multi vector, dense embedding, sparse embedding
    Image Search with MilvusSemantic Searchvector search, dynamic field
    Multimodal Search using Multi VectorsSemantic Searchmulti vector, hybrid search
    Movie Recommendation with MilvusRecommendation Systemvector search
    Graph RAG with MilvusRAGgraph search
    Contextual Retrieval with MilvusQuickstartvector search
    Vector VisualizationQuickstartvector search
    HDBSCAN Clustering with MilvusQuickstartvector search
    Use ColPali for Multi-Modal Retrieval with MilvusQuickstartvector search

    Ecosystem and Integration

    Milvus integrates with a comprehensive suite of AI development tools, such as LangChain, LlamaIndex, OpenAI and HuggingFace, making it an ideal vector store for GenAI applications such as Retrieval-Augmented Generation (RAG). Milvus works with both open-source embedding models and embedding services, in text, image and video modalities. Milvus also provides a convenient utility pymilvus[model], users can use the simple wrapper code to transform unstructured data into vector embeddings and leverage reranking models for optimized search results. The Milvus ecosystem also includes Attu for GUI-based administration, Birdwatcher for system debugging, Prometheus/Grafana for monitoring, Milvus CDC for data synchronization, VTS for data migration and data connectors for Spark, Kafka, Fivetran, and Airbyte to build search pipelines.

    Check out https://milvus.io/docs/integrations_overview.md for more details.

    Documentation

    For guidance on installation, usage, deployment, and administration, check out Milvus Docs. For technical milestones and enhancement proposals, check out issues on GitHub.

    Contributing

    The Milvus open-source project accepts contributions from everyone. See Guidelines for Contributing for details on submitting patches and the development workflow. See our community repository to learn about project governance and access more community resources.

    Build Milvus from Source Code

    Requirements:

    • Linux systems (Ubuntu 20.04 or later recommended):

      Go: >= 1.21
CMake: >= 3.26.4 && CMake < 4
GCC: >= 11
Python: > 3.8 and  <= 3.11

    • MacOS systems with x86_64 (Big Sur 11.5 or later recommended):

      Go: >= 1.21
CMake: >= 3.26.4 && CMake < 4
llvm: >= 15
Python: > 3.8 and  <= 3.11

    • MacOS systems with Apple Silicon (Monterey 12.0.1 or later recommended):

      Go: >= 1.21 (Arch=ARM64)
CMake: >= 3.26.4 && CMake < 4
llvm: >= 15
Python: > 3.8 and  <= 3.11

    Clone Milvus repo and build.

    # Clone github repository.
$ git clone https://github.com/milvus-io/milvus.git

# Install third-party dependencies.
$ cd milvus/
$ ./scripts/install_deps.sh

# Compile Milvus.
$ make

    For full instructions, see developer's documentation.

    Community

    Join the Milvus community on Discord to share your suggestions, advice, and questions with our engineering team.

    To learn the latest news about Milvus, follow us on social media:

    You can also check out our FAQ page to discover solutions or answers to your issues or questions, and subscribe to Milvus mailing lists:

    Reference

    Reference to cite when you use Milvus in a research paper:

    @inproceedings{2021milvus,
  title={Milvus: A Purpose-Built Vector Data Management System},
  author={Wang, Jianguo and Yi, Xiaomeng and Guo, Rentong and Jin, Hai and Xu, Peng and Li, Shengjun and Wang, Xiangyu and Guo, Xiangzhou and Li, Chengming and Xu, Xiaohai and others},
  booktitle={Proceedings of the 2021 International Conference on Management of Data},
  pages={2614--2627},
  year={2021}
}

@article{2022manu,
  title={Manu: a cloud native vector database management system},
  author={Guo, Rentong and Luan, Xiaofan and Xiang, Long and Yan, Xiao and Yi, Xiaomeng and Luo, Jigao and Cheng, Qianya and Xu, Weizhi and Luo, Jiarui and Liu, Frank and others},
  journal={Proceedings of the VLDB Endowment},
  volume={15},
  number={12},
  pages={3548--3561},
  year={2022},
  publisher={VLDB Endowment}
}

    <img src="https://avata

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    faiss
    facebookresearch·C++

    面向稠密向量的高效相似度检索与聚类库，支持十亿级向量近邻搜索，是构建向量数据库和 RAG 检索层的底层引擎。

    向量检索+4向量数据库
    41344+534.6kv1.15.12026-11-10 21:17 UTC
    qdrant
    qdrant·Rust

    高性能大规模向量数据库与向量检索引擎，用 Rust 编写，支持 HNSW、混合检索和图像搜索，适合为 AI 应用和 RAG 提供相似度检索能力。

    向量数据库+4向量数据库
    35503+392.8kv1.20.02026-11-10 21:32 UTC
    dragonfly
    dragonflydb·C++

    兼容 Redis 和 Memcached 协议的内存数据库，用多线程架构提升吞吐量，可直接替换现有缓存层。

    内存数据库+4向量数据库
    31860+131.3kv2.12026-11-10 16:51 UTC
    chroma
    chroma-core·Rust

    开源的向量数据库，专门给 AI 应用做检索基础设施，支持 RAG、语义搜索和智能体记忆存储，底层用 Rust 重写后性能更强。

    向量数据库+4向量数据库
    29672+182.6k1.5.92026-11-07 04:52 UTC