Milvus: Distributed Vector Database for Billion-Scale Search

Quick Start: Docker Standalone & Collection Schema

Run Milvus standalone locally with Docker, then define a typed collection schema.

# Start Milvus standalone with Docker Compose
# curl -sfL https://raw.githubusercontent.com/milvus-io/milvus/master/scripts/standalone_embed.sh | bash

pip install pymilvus openai

from pymilvus import (
    MilvusClient, DataType, Collection,
    FieldSchema, CollectionSchema
)

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

# Define schema — primary key + vector field + scalar metadata
schema = MilvusClient.create_schema(auto_id=False, enable_dynamic_field=True)
schema.add_field(field_name="id",       datatype=DataType.INT64,      is_primary=True)
schema.add_field(field_name="vector",   datatype=DataType.FLOAT_VECTOR, dim=1536)
schema.add_field(field_name="text",     datatype=DataType.VARCHAR,    max_length=2048)
schema.add_field(field_name="category", datatype=DataType.VARCHAR,    max_length=64)
schema.add_field(field_name="year",     datatype=DataType.INT32)

# Create collection
client.create_collection(
    collection_name="articles",
    schema=schema,
    consistency_level="Strong"    # Strong | Bounded | Eventually
)
print(client.describe_collection("articles"))

Create HNSW Index, Insert Vectors & Search

Build an HNSW index on the vector field, insert data in batches, then run filtered ANN search.

import openai

oa = openai.OpenAI()

# Build HNSW index before inserting large datasets
index_params = client.prepare_index_params()
index_params.add_index(
    field_name="vector",
    index_type="HNSW",           # HNSW | IVF_FLAT | IVF_PQ | DiskANN | SCANN
    metric_type="COSINE",        # COSINE | L2 | IP
    params={"M": 16, "efConstruction": 200}
)
client.create_index(collection_name="articles", index_params=index_params)
client.load_collection("articles")  # load into memory for search

# Generate embeddings and insert
docs = [
    {"id": 1, "text": "Milvus scales to billion vectors", "category": "db",  "year": 2024},
    {"id": 2, "text": "HNSW provides high recall ANN search", "category": "ai", "year": 2024},
    {"id": 3, "text": "DiskANN enables disk-based billion-scale search", "category": "db","year": 2023},
]
embs = oa.embeddings.create(model="text-embedding-3-small", input=[d["text"] for d in docs])
for d, e in zip(docs, embs.data):
    d["vector"] = e.embedding

client.insert(collection_name="articles", data=docs)
client.flush(collection_name="articles")   # ensure data is persisted

# Filtered ANN search
query_vec = oa.embeddings.create(
    model="text-embedding-3-small", input=["distributed vector search"]
).data[0].embedding

results = client.search(
    collection_name="articles",
    data=[query_vec],
    filter='category == "db" and year >= 2024',   # scalar filter
    limit=3,
    output_fields=["text", "category", "year"]
)
for hits in results:
    for hit in hits:
        print(f"score={hit['distance']:.4f} | {hit['entity']['text']}")

DiskANN: Billion-Scale with Disk Storage

When your collection exceeds available RAM, DiskANN stores the index on SSD while maintaining good recall and latency.

# DiskANN index — ideal for 100M+ vectors exceeding RAM
index_params = client.prepare_index_params()
index_params.add_index(
    field_name="vector",
    index_type="DISKANN",
    metric_type="COSINE",
    params={
        "search_list_size": 100,    # higher = better recall, slower build
        "max_degree": 56,           # graph degree
        "pq_code_budget_gb": 2.0,   # budget for PQ codes in RAM
        "build_dram_budget_gb": 32.0 # RAM for build phase
    }
)

# Zilliz Cloud connection (API-compatible with PyMilvus)
cloud_client = MilvusClient(
    uri="https://YOUR_CLUSTER.api.gcp-us-west1.zillizcloud.com",
    token="YOUR_ZILLIZ_API_KEY"
)
# Same SDK — just change the connection URI

Milvus vs Qdrant vs Weaviate vs Pinecone