Milvus Vector Database Skill

Operate Milvus vector databases directly through Python code using the pymilvus SDK. This skill covers the full lifecycle — connecting, schema design, collection management, vector CRUD, search, hybrid search, indexing, partitions, databases, and RBAC.

When to Use

Use this skill when the user wants to:

• - Connect to a Milvus instance (local, standalone, cluster, or Milvus Lite)
• Create collections with custom schemas
• Insert, upsert, search, query, get, or delete vectors
• Perform hybrid search with reranking
• Manage indexes, partitions, databases
• Set up users, roles, and access control (RBAC)
• Build RAG pipelines, semantic search, or recommendation systems with Milvus

Requirements

• - Python 3.8+
• INLINECODE1 (pip install pymilvus)
• A running Milvus instance, or use Milvus Lite (embedded, file-based) for development

Capabilities Overview

Area	What You Can Do
Connection	Connect to Milvus Lite, Standalone, Cluster, or Zilliz Cloud
Collections

Create (quick or custom schema), list, describe, drop, rename, load, release | | Vectors | Insert, upsert, search, hybrid search, query, get, delete | | Indexes | Create (AUTOINDEX, HNSW, IVF_FLAT, etc.), list, describe, drop | | Partitions | Create, list, load, release, drop | | Databases | Create, list, switch, drop | | RBAC | Users, roles, privileges management |

---

Connection

CODEBLOCK0

Parameters:

Parameter	Type	Description
INLINECODE3	str	INLINECODE4 for Milvus Lite, "http://host:19530" for server
INLINECODE6

str | API key or "username:password" |
| user | str | Username (alternative to token) |
| password | str | Password (alternative to token) |
| db_name | str | Target database (default: "default") |
| timeout | float | Operation timeout in seconds |

Async Client

CODEBLOCK1

---

Collection Management

Quick Create (auto schema + auto index + auto load)

CODEBLOCK2

This automatically creates:

• - id field (INT64, primary key, autoid)
• INLINECODE14 field (FLOATVECTOR, dim=dimension)
• AUTOINDEX on vector field
• Collection is auto-loaded

Custom Schema Create

CODEBLOCK3

Supported Data Types

Scalar types:

DataType	Notes
INLINECODE15	Boolean
INLINECODE16 / INT16 / INT32 / INLINECODE19

Integers |
| DataType.FLOAT / DOUBLE | Floating point |
| DataType.VARCHAR | String (requires max_length) |
| DataType.JSON | JSON object |
| DataType.ARRAY | Array (requires element_type, max_capacity) |

Vector types:

DataType	Notes
INLINECODE28	Float32 vector (requires dim)
INLINECODE30

Float16 vector (requires dim) |
| DataType.BFLOAT16_VECTOR | BFloat16 vector (requires dim) |
| DataType.BINARY_VECTOR | Binary vector (requires dim) |
| DataType.SPARSE_FLOAT_VECTOR | Sparse vector (no dim needed) |

add_field Parameters

CODEBLOCK4

Other Collection Operations

CODEBLOCK5

Collection Guidance

• - Quick create is best for prototyping; use custom schema for production.
• A collection must be loaded before search or query operations.
• Before dropping a collection, confirm with the user — this deletes all data.
• Use enable_dynamic_field=True to allow inserting fields not defined in the schema.

---

Vector Operations

Target collection must exist and be loaded.

Insert

CODEBLOCK6

Upsert (insert or update if PK exists)

CODEBLOCK7

Search (vector similarity)

CODEBLOCK8

Hybrid Search (multi-vector with reranking)

CODEBLOCK9

Query (filter-based retrieval)

CODEBLOCK10

Get (by primary key)

CODEBLOCK11

Delete

CODEBLOCK12

Filter Expression Syntax

Expression	Example
Comparison	INLINECODE39
Equality

status == "active" | | IN list | id in [1, 2, 3] | | AND/OR | age > 20 and status == "active" | | String match | text like "hello%" | | Array contains | ARRAY_CONTAINS(tags, "ml") | | JSON field | json_field["key"] > 100 | | Match all | id > 0 |

Vector Guidance

• - The data parameter in search must match the collection's vector dimension exactly.
• For text-to-vector search, convert text to vectors using an embedding model first.
• For large inserts, batch data into chunks (e.g., 1000 rows per batch).
• Always specify output_fields to control which fields are returned.

---

Index Management

Create Index

CODEBLOCK13

Common Index Types

Index Type	For	Key Params	Notes
INLINECODE49	Dense vectors	Auto-tuned	Recommended for most cases
INLINECODE50

Dense vectors | None | Brute force, 100% recall | | IVF_FLAT | Dense vectors | nlist | Good balance | | IVF_SQ8 | Dense vectors | nlist | Compressed, less memory | | HNSW | Dense vectors | M, efConstruction | High recall, more memory | | DISKANN | Dense vectors | None | Disk-based, large datasets | | SPARSE_INVERTED_INDEX | Sparse vectors | drop_ratio_build | For sparse vectors | | SPARSE_WAND | Sparse vectors | drop_ratio_build | Faster sparse search |

Metric Types

Metric	Description	Use With
INLINECODE63	Cosine similarity (larger = more similar)	Dense vectors
INLINECODE64

Euclidean distance (smaller = more similar) | Dense vectors | | "IP" | Inner product (larger = more similar) | Dense & Sparse vectors |

Other Index Operations

CODEBLOCK14

Index Guidance

• - AUTOINDEX is recommended for most use cases.
• An index is required before loading a collection.
• After creating an index, load the collection before searching.
• Sparse vectors only support "IP" metric type.

---

Partition Management

CODEBLOCK15

Partition Guidance

• - Every collection has a _default partition.
• Use is_partition_key=True on a field to enable automatic partitioning by field value.
• A partition must be loaded before search.
• Before dropping a partition, confirm with the user — all data in it will be deleted.

---

Database Management

CODEBLOCK16

Database Guidance

• - Every Milvus instance has a "default" database.
• Before dropping a database, all collections in it must be dropped first.

---

User & Role Management (RBAC)

User Operations

CODEBLOCK17

Role Operations

CODEBLOCK18

Built-in Privilege Groups

Group	Scope
INLINECODE71	Full cluster access
INLINECODE72

Read-only cluster access | | ClusterReadWrite | Read-write cluster access | | DatabaseAdmin | Full database access | | DatabaseReadOnly | Read-only database access | | DatabaseReadWrite | Read-write database access | | CollectionAdmin | Full collection access | | CollectionReadOnly | Read-only collection access | | CollectionReadWrite | Read-write collection access |

Common Individual Privileges

INLINECODE80, Query, Insert, Delete, Upsert, CreateIndex, DropIndex, CreateCollection, DropCollection, Load, Release, CreatePartition, INLINECODE92

RBAC Guidance

• - Recommended workflow: create role → grant privileges → create user → assign role.
• Use "*" for collectionname/dbname to grant on all resources.
• Before dropping a user or role, confirm with the user.

---

Common Patterns

RAG Pipeline Pattern

CODEBLOCK19

Quick Semantic Search Pattern

CODEBLOCK20

---

General Guidance

• - Always check if pymilvus is installed: pip install pymilvus.
• For quick prototyping, use Milvus Lite (uri="./file.db") — no server needed.
• A collection must be loaded into memory before search/query.
• The vector dimension in search data must exactly match the collection schema.
• For text queries, users need an embedding model to convert text to vectors first. Suggest pymilvus[model] for built-in embedding support.
• Before any destructive operation (drop collection, drop database, delete vectors), always confirm with the user.
• Use enable_dynamic_field=True when the schema may evolve.
• For large-scale inserts, batch data into chunks of ~1000 rows.
• Prefer AUTOINDEX unless the user has specific performance requirements.