登录

pymilvus

milvus-ioApache-2.0活跃维护

Milvus 向量数据库的官方 Python SDK，让 Python 开发者能直接建集合、写入向量并做相似度检索，是搭建 RAG 和向量搜索应用的基础客户端。

Python SDK for Milvus Vector Database

v2.6.18发布于 2026-11-06 08:23 UTC1,425star +2 / 7 天496fork同步于 2026-11-10 23:49 UTC
向量数据库Python SDK相似度检索MilvusRAG
对比 GitHubannsfaissfaiss-vector-databasemilvusmilvus-lite

怎么上手

由 README 提炼，执行前请确认
原文 
pip3 install pymilvus
硬件要求按 README 说明

README 未说明；本包是客户端 SDK，需另行部署 Milvus 服务端，硬件要求取决于服务端

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

Milvus 向量数据库的官方 Python SDK，用来在 Python 里建集合、写入和检索向量

Milvus 向量数据库的官方 Python SDK，让 Python 开发者能直接建集合、写入向量并做相似度检索，是搭建 RAG 和向量搜索应用的基础客户端。

核心能力

  • 提供 pip 安装的 Python 客户端，连接 Milvus 服务端
  • 按 Milvus 版本匹配对应 PyMilvus 版本，如 2.6.* 对应 2.6.X
  • 可选 model 扩展，安装 pymilvus[model] 获得 milvus-model
  • 可选 bulk_writer 扩展，安装 pymilvus[bulk_writer] 支持批量写入
  • 支持从 milvus-proto 生成 Python 协议文件

容易踩的坑

PyMilvus 版本需与 Milvus 服务端版本对应，装错大版本可能不兼容；zsh 下安装 pymilvus[model] 需加引号写成 pip install "pymilvus[model]"；从源码开发需先 git submodule update --init 拉取 proto 子模块，且本地开发命令依赖 uv

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v2.6.18
全部 release
  1. v2.6.18PyMilvus v2.6.18 Release Notes最新正式版2026-11-06 · 2026-11-06 08:23 UTC

    PyMilvus 2.6.18 targets the Milvus 2.6 series and requires Python 3.8 or later.

    What's Changed

    Features and enhancements

    • Add row-level security (RLS) policy and principal-tag management to MilvusClient and AsyncMilvusClient. Supported read and write operations accept rls_principal for server-side enforcement. These APIs require a deployment with RLS support. #3853
    • Add configurable Volume upload policies, with paginated listings and bounded concurrency for large batches. VolumeFileManager.upload_file_to_volume() now skips an existing object by default when its key and byte size match. Set upload_policy=UploadPolicy.OVERWRITE to replace matching objects. #3805
    • Accept custom HTTP headers in bulk_import(), get_import_progress(), and list_import_jobs(). The SDK also exports IDEMPOTENCY_KEY_HEADER for use with servers that support idempotent bulk imports. #3808

    Performance and bug fixes

    • Restore linear-time decoding of nullable scalar output fields such as VARCHAR in search results. #3724
    • Implement the async RPC handlers for AsyncMilvusClient.flush_all(), get_compaction_plans(), and update_replicate_configuration(). #3891
    • Preserve existing application log handlers when importing PyMilvus and scope BulkWriter logging to the pymilvus.bulk_writer logger tree. #3847 · #3849
    • Return None in QueryIterator.get_cursor().str_pk before the first page when the primary key is VARCHAR. #3860
    • Raise ParamError for empty binary-vector lists in batch writes and reject out-of-range ports during connection validation. #3697 · #3833

    Documentation, tests, and CI

    • Standardize non-ORM docstrings on Google style. #3748
    • Expand unit test coverage, consolidate overlapping tests, and ensure Windows CI reports failures from either the unit or Lite integration suite. #3879 · #3832

    Full Changelog

  2. v3.0.2PyMilvus v3.0.2 Release Notes2026-09-18 · 2026-09-18 08:53 UTC
    查看更新说明

    What's Changed

    Full Changelog: https://github.com/milvus-io/pymilvus/compare/v3.0.1...v3.0.2

  3. v3.0.1PyMilvus v3.0.1 Release Notes2026-07-29 · 2026-07-29 14:55 UTC
    查看更新说明

    PyMilvus v3.0.1 Release Notes

    PyMilvus 3.0.1 is the recommended Python SDK release for Milvus 3.0.

    Compatibility: PyMilvus 3.0.1 requires Python 3.9 or later. Python 3.8 is no longer supported.

    What's Changed

    New APIs and capabilities

    • Add DataType.TEXT handling across schema validation, result decoding, and BulkWriter paths (#3473).
    • Add get_replicate_info() to MilvusClient and AsyncMilvusClient (#3516).
    • Add sync and async dump_messages() support for data-salvage workflows (#3575).
    • Add role and user descriptions, including alter_role() and role-description output from describe_role() (#3608).
    • Add search aggregation and correctly forward group_by_fields (#3647).
    • Add FunctionChain APIs, forward every specified stage during search, and add the XGBoost expression helper (#3652, #3662, #3685).

    Schema evolution and function fields

    • Add alter_collection_schema() and support dynamically adding struct fields (#3504, #3569).
    • Route collection-schema mutations through AlterCollectionSchema, retain a legacy AddCollectionField fallback only for servers returning UNIMPLEMENTED, and keep drop-function RPC behavior compatible (#3572, #3706).
    • Support external field mappings and add-field fallback for external collections (#3584, #3728).
    • Add drop_function_field(), support MinHash function fields, and require bound index parameters when adding a function field (#3600, #3630, #3672).
    • Keep primary-key, partition-key, and clustering-key metadata synchronized when fields are added (#3667).

    Iterators, search results, and type handling

    • Move query, Search Iterator V1, and Search Iterator V2 implementations to the client layer while preserving ORM compatibility and aligning ORM deprecation warnings (#3707, #3709, #3714, #3718).
    • Preserve external filtering when creating search iterators (#3730).
    • Add shared field-data and nested-value decoders, remove the dead V1 row extractor, and avoid quadratic nullable-scalar result decoding (#3552, #3555, #3562, #3605, #3713, #3725).
    • Consolidate logical type metadata across search, insert, array, row, and vector paths to remove inconsistent type-specific behavior (#3474, #3481, #3499, #3539, [#3548](https://github.com/milvus-io/pymil
  4. v2.6.17PyMilvus v2.6.17 Release Notes2026-07-17 · 2026-07-17 06:23 UTC
    查看更新说明

    What's Changed

    Protocol and compatibility

    • Sync milvus-proto with Milvus v2.6.20 and regenerate Python protobuf bindings, including support for the inclusive DumpMessages start option (#3687).

    Bug fixes

    • Materialize array query results as standard Python lists (source #3653; 2.6 #3654).
    • Preserve struct-array fields when rebuilding schemas from dictionaries, and support partial updates for struct-array fields (sources #3651 and #3671; 2.6 #3656 and #3675).
    • Support struct vector fields in BulkWriter and preserve remote upload path prefixes on Windows (sources #3631 and #3658; 2.6 #3669).
    • Keep primary-key, partition-key, and clustering-key metadata synchronized when fields are added to a schema (source #3649; 2.6 #3668).
    • Allow null values in an auto_id primary-key DataFrame column during insert (source #3673; 2.6 #3677).
    • Correctly unpack async schema results before running collection optimization (source #3681; 2.6 #3682).
    • Reject boolean values used as sparse-vector indices (source #3638; 2.6 #3664; fixes #3622).
    • Require target_size in compact() to be a positive integer (source #3642; 2.6 #3643; fixes #3169).

    Build and release tooling

    • Isolate protobuf generation dependencies from normal development dependencies (source #3689; 2.6 #3691).
    • Keep 2.6 development-version generation on vcs-versioning<2 semantics (source #3633; 2.6 #3635).

    Full Changelog: https://github.com/milvus-io/pymilvus/compare/v2.6.16...v2.6.17

  5. v2.6.16PyMilvus v2.6.16 Release Notes2026-06-25 · 2026-06-25 02:10 UTC
    查看更新说明

    What's Changed

    Features and enhancements

    • Add element-level search support (source #3208; 2.6 #3491).
    • Add dump_messages support for data salvage (source #3573; 2.6 #3574).
    • Add role and user description support in auth APIs (sources #3601 and #3602; 2.6 #3609).
    • Improve volume upload interaction, including concurrency and progress behavior (source #3619; 2.6 #3628).

    Bug fixes

    • Harden sync and async connection recovery by protecting reconnect channel swaps and validating channel reclamation (sources #3542, #3564, and #3579; 2.6 #3566, #3571, and #3582).
    • Preserve connection timeout settings during recovery (source #3610; 2.6 #3614).
    • Map only connection-setup RPC failures to CONNECT_FAILED (source #3589; 2.6 #3591).
    • Reject boolean values in sparse vector validation and improve the related error message (source #3624; 2.6 #3625).

    Type handling and client internals

    • Continue the logical type model migration across array helpers, row insert and extraction, field-data result and nested decoders, vector predicates, and vector-array dtype facts (sources #3518, #3501, #3536, #3537, #3554, #3604, and #3606; 2.6 #3549, #3551, #3553, #3556, #3558, #3607, and #3612).
    • Remove the dead V1 row extractor (source #3561; 2.6 #3563).

    Documentation and tests

    • Update async client examples (source #3597; 2.6 #3598).
    • Classify maintained test suites (source #3592; 2.6 #3596).

    Full Changelog: https://github.com/milvus-io/pymilvus/compare/v2.6.15...v2.6.16

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

    Milvus Python SDK

    version Supported Python Versions Downloads Downloads Downloads

    license Static Badge

    Python SDK for Milvus. To contribute code to this project, please read our contribution guidelines first. If you have some ideas or encounter a problem, you can find us in the Slack channel #py-milvus.

    Compatibility

    The following collection shows Milvus versions and recommended PyMilvus versions:

    Milvus versionRecommended PyMilvus version
    1.0.*1.0.1
    1.1.*1.1.2
    2.0.*2.0.2
    2.1.*2.1.3
    2.2.*2.2.15
    2.3.*2.3.7
    2.4.*2.4.X
    2.5.*2.5.X
    2.6.*2.6.X
    3.0.*3.0.X

    Installation

    You can install PyMilvus via pip or pip3 for Python 3.9+:

    $ pip3 install pymilvus
$ pip3 install pymilvus[model] # for milvus-model
$ pip3 install pymilvus[bulk_writer] # for bulk_writer

    You can install a specific version of PyMilvus by:

    $ pip3 install pymilvus==2.4.10

    You can upgrade PyMilvus to the latest version by:

    $ pip3 install --upgrade pymilvus

    FAQ

    Local development commands use uv. Install uv before running the make targets below.

    Q1. How to get submodules?

    A1. The following command will get the protos matching to the generated files, for protos of certain version, see milvus-proto for details.

    $ git submodule update --init

    Q2. How to generate python files from milvus-proto?

    A2.

    $ make gen_proto

    Q3. How to use the local PyMilvus repository for Milvus server?

    A3.

    $ make install

    Q4. How to check and auto-fix the coding styles?

    A4.

    make lint
make format

    Q5. How to set up pre-commit hooks to automatically check and fix the coding styles?

    Once installed, the hooks will automatically run make format and make lint before each commit. If the checks fail, the commit will be aborted, and you'll need to fix the issues before committing again.

    A5. Pre-commit hooks help ensure code quality by automatically running linting and formatting checks before each commit.

    # Install pre-commit (if not already installed)
$ uv tool install pre-commit

# Install the git hook scripts
$ pre-commit install

    Q7. How to run the maintained test suites?

    A7

    $ uv sync --group dev
$ make unittest
$ make integration-lite

    Q8. zsh: no matches found: pymilvus[model], how do I solve this?

    A8

    $ pip install "pymilvus[model]"

    Documentation

    Documentation is available online: https://milvus.io/api-reference/pymilvus/v2.6.x/About.md

    Developing package releases

    The commits on the development branch of each version will be packaged and uploaded to Test PyPI.

    The package name generated by the development branch is x.y.z.rc, where is the number of commits that differ from the most recent release.

    • For example, after the release of 2.3.4, two commits were submitted on the 2.3 branch. The version number of the latest commit of 2.3 branch is 2.3.5.rc2.

    • For example, after the release of 2.3.4, 10 commits were submitted on the master branch. The version number of the latest commit of master branch is 2.4.0.rc10.

    To install the package on Test PyPi, you need to append --extra-index-url after pip, for example:

    $ python3 -m pip install --extra-index-url https://test.pypi.org/simple/ pymilvus==2.1.0.dev66

    License

    Apache License 2.0

    相关项目

    同分类里 star 最高的项目
    对比全部 →
    milvus
    milvus-io·Go

    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。

    向量数据库+4向量数据库
    46632+204.3kv3.0.52026-11-10 18:36 UTC
    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+834.6kv1.15.12026-11-10 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+141.3kv2.12026-11-10 16:51 UTC