title: Milvus description: High-performance vector database — Milvus for RAG: collections, indexing algorithms, partitioning, GPU acceleration, and Zilliz Cloud managed service. category: 09-RAG tags: [milvus, vector-store, high-performance, gpu, zilliz, rag] status: ready last_updated: 2026-09-06

Milvus¶

Milvus is a high-performance vector database built for scale. It supports multiple indexing algorithms (IVF, HNSW, ANNOY), GPU acceleration, and handles billions of vectors. Use Milvus when you need maximum performance at scale — or use Zilliz Cloud for a managed version.

Installation / Setup¶

pip install pymilvus                    # or: uv add pymilvus
# docker: docker compose up -d (for local Milvus)
# cloud: https://zilliz.com/cloud/

Core Syntax¶

Client setup¶

from pymilvus import MilvusClient

# local (Milvus Lite — in-process)
client = MilvusClient("milvus.db")

# local (Milvus server)
client = MilvusClient(uri="http://localhost:19530")

# Zilliz Cloud
client = MilvusClient(uri="https://xxx.zillizcloud.com:19530", token="YOUR_TOKEN")

Create collection¶

from pymilvus import DataType

schema = client.create_schema(auto_id=True, enable_dynamic_field=True)
schema.add_field("id", DataType.INT64, is_primary=True)
schema.add_field("vector", DataType.FLOAT_VECTOR, dim=1536)
schema.add_field("text", DataType.VARCHAR, max_length=2000)
schema.add_field("source", DataType.VARCHAR, max_length=200)

client.create_collection(
    collection_name="documents",
    schema=schema,
)

Insert data¶

client.insert(collection_name="documents", data=[
    {"vector": [0.1]*1536, "text": "Python is great for ML", "source": "tutorial"},
    {"vector": [0.2]*1536, "text": "JavaScript runs in browsers", "source": "web"},
])

Search¶

results = client.search(
    collection_name="search",
    data=[[0.1]*1536],     # query vectors (supports batch)
    limit=5,
    output_fields=["text", "source"],
)
for hits in results:
    for hit in hits:
        print(f"Score: {hit['distance']:.3f}, Text: {hit['entity']['text']}")

LangChain + Milvus¶

from langchain_community.vectorstores import Milvus
from langchain_openai import OpenAIEmbeddings

vectorstore = Milvus.from_documents(
    chunks, OpenAIEmbeddings(),
    connection_args={"host": "localhost", "port": "19530"},
    collection_name="documents",
)

docs = vectorstore.similarity_search("query", k=5)

Common Operations¶

Index creation¶

from pymilvus import CollectionSchema

# HNSW index (fast, good quality)
client.create_index(
    collection_name="documents",
    field_name="vector",
    index_params={
        "index_type": "HNSW",
        "metric_type": "COSINE",
        "params": {"M": 16, "efConstruction": 256},
    },
)

# IVF_FLAT index (good for large datasets)
client.create_index(
    collection_name="documents",
    field_name="vector",
    index_params={
        "index_type": "IVF_FLAT",
        "metric_type": "COSINE",
        "params": {"nlist": 128},
    },
)

Search with filtering¶

results = client.search(
    collection_name="documents",
    data=[[0.1]*1536],
    limit=5,
    filter='source == "report" and year > 2024',
    output_fields=["text", "source", "year"],
)

Collection management¶

# list collections
collections = client.list_collections()

# get collection info
info = client.describe_collection("documents")

# delete collection
client.drop_collection("documents")

Partitioning (for large datasets)¶

# create partition
client.create_partition("documents", partition_name="2024")

# insert into partition
client.insert("documents", data=[...], partition_name="2024")

# search in partition
results = client.search("documents", data=[[0.1]*1536], partition_names=["2024"])

Practical Examples¶

Example 1 — Milvus RAG pipeline¶

from pymilvus import MilvusClient
from langchain_community.vectorstores import Milvus
from langchain_openai import OpenAIEmbeddings, ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnablePassthrough

client = MilvusClient(uri="http://localhost:19530")
vectorstore = Milvus(client, OpenAIEmbeddings(), collection_name="documents")
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})

rag = (
    {"context": retriever | (lambda d: "\n\n".join(x.page_content for x in d)),
     "question": RunnablePassthrough()}
    | ChatPromptTemplate.from_template("Answer:\n{context}\n\nQ: {question}")
    | ChatOpenAI(model="gpt-4o")
    | StrOutputParser()
)

Common Patterns¶

Pattern: index selection¶

Dataset size	Recommended index	Why
<100K	HNSW	Fast, high quality
100K–10M	IVF_FLAT	Good balance
>10M	IVF_PQ	Memory efficient
GPU available	GPU_IVF_FLAT	Fastest for large scale

Best Practices¶

Use Milvus Lite (milvus.db) for local development.
Use Zilliz Cloud for production — fully managed.
Choose index based on dataset size — HNSW for small, IVF for large.
Use partitions to organize large datasets by date/category.
Milvus supports GPU acceleration — use it for very large scale.

Common Mistakes¶

Mistake	Fix
Using Milvus server for local dev	Use Milvus Lite (`milvus.db`)
Not creating an index	Always create an index before search
Wrong metric type	Match to embedding model: COSINE for normalized

Troubleshooting¶

Symptom	Cause / Fix
Connection refused	Milvus server not running
`Collection not found`	Create collection or check name
Slow queries	Create proper index; check dataset size
OOM	Use IVF_PQ for compression; increase memory

qdrant.md — production alternative
pinecone.md — managed alternative
faiss.md — local alternative
rag.md — the complete pipeline

Official Documentation¶

Milvus docs — https://milvus.io/docs/
Zilliz Cloud — https://zilliz.com/cloud/
PyMilvus — <https://milvus.io/docs/install-pymilvus.md

Last Updated¶

2026-09-06 — Initial version.