openova/platform/milvus

Milvus

Vector database for similarity search. Application Blueprint (see docs/PLATFORM-TECH-STACK.md §4.6). Backbone of RAG pipelines in bp-cortex (paired with BGE embeddings).

Status: Accepted | Updated: 2026-04-27

---

Overview

Milvus provides high-performance vector similarity search for RAG, recommendation systems, and AI applications.

flowchart TB
    subgraph Milvus["Milvus Cluster"]
        Proxy[Proxy]
        QueryNode[Query Nodes]
        DataNode[Data Nodes]
        IndexNode[Index Nodes]
    end

    subgraph Storage["Storage"]
        etcd[etcd<br/>Metadata]
        SeaweedFS[SeaweedFS<br/>Object Storage]
    end

    App[Application] --> Proxy
    Proxy --> QueryNode
    Proxy --> DataNode
    DataNode --> SeaweedFS
    QueryNode --> SeaweedFS
    Proxy --> etcd

---

Why Milvus?

Feature	Benefit
Billion-scale vectors	Enterprise-grade scalability
Hybrid search	Dense + sparse vectors
Multiple indexes	IVF, HNSW, DiskANN
GPU acceleration	Optional GPU indexing
Cloud-native	Kubernetes-native deployment

---

Use Cases

Use Case	Description
RAG	Document chunk retrieval
Semantic search	Natural language queries
Recommendation	Similar item retrieval
Image search	Visual similarity
Anomaly detection	Outlier identification

---

Configuration

Helm Values

cluster:
  enabled: true

etcd:
  replicaCount: 3
  persistence:
    size: 10Gi

seaweedfs:
  enabled: false  # Use external SeaweedFS
  externalS3:
    enabled: true
    host: seaweedfs.storage.svc
    port: 9000
    accessKey: ""  # From ESO
    secretKey: ""  # From ESO
    bucketName: milvus

proxy:
  replicas: 2
  resources:
    requests:
      cpu: 500m
      memory: 1Gi

queryNode:
  replicas: 2
  resources:
    requests:
      cpu: 1
      memory: 4Gi

dataNode:
  replicas: 2
  resources:
    requests:
      cpu: 500m
      memory: 2Gi

indexNode:
  replicas: 1
  resources:
    requests:
      cpu: 1
      memory: 4Gi

---

Collection Schema

from pymilvus import Collection, FieldSchema, CollectionSchema, DataType

fields = [
    FieldSchema(name="id", dtype=DataType.VARCHAR, max_length=64, is_primary=True),
    FieldSchema(name="document_id", dtype=DataType.VARCHAR, max_length=64),
    FieldSchema(name="chunk_index", dtype=DataType.INT64),
    FieldSchema(name="text", dtype=DataType.VARCHAR, max_length=65535),
    FieldSchema(name="source", dtype=DataType.VARCHAR, max_length=32),
    FieldSchema(name="dense_vector", dtype=DataType.FLOAT_VECTOR, dim=1024),
    FieldSchema(name="sparse_vector", dtype=DataType.SPARSE_FLOAT_VECTOR),
]

schema = CollectionSchema(fields, description="Document chunks")
collection = Collection("documents", schema)

---

Index Types

Index	Use Case	Memory
HNSW	High recall, fast query	High
IVF_FLAT	Balanced	Medium
IVF_SQ8	Memory-efficient	Low
DiskANN	Billion-scale	Disk-based
GPU_IVF_FLAT	GPU-accelerated	GPU memory

Create Index

index_params = {
    "metric_type": "COSINE",
    "index_type": "HNSW",
    "params": {"M": 16, "efConstruction": 256}
}
collection.create_index("dense_vector", index_params)

---

Hybrid Search

Combine dense and sparse vectors:

from pymilvus import AnnSearchRequest, WeightedRanker

# Dense search
dense_req = AnnSearchRequest(
    data=[dense_vector],
    anns_field="dense_vector",
    param={"metric_type": "COSINE", "params": {"ef": 64}},
    limit=20
)

# Sparse search
sparse_req = AnnSearchRequest(
    data=[sparse_vector],
    anns_field="sparse_vector",
    param={"metric_type": "IP"},
    limit=20
)

# Combine with weighted ranker
results = collection.hybrid_search(
    [dense_req, sparse_req],
    rerank=WeightedRanker(0.7, 0.3),
    limit=10
)

---

Partition Strategy

# Partition by source for isolation
collection.create_partition("compliance")
collection.create_partition("infrastructure")
collection.create_partition("ephemeral")

# Query specific partition
results = collection.search(
    data=[query_vector],
    anns_field="dense_vector",
    partition_names=["compliance"],
    limit=10
)

---

Monitoring

Metric	Query
Query latency	milvus_proxy_search_latency
Insert rate	milvus_datanode_flush_buffer_op_total
Memory usage	milvus_querynode_memory_usage
Collection size	milvus_datacoord_stored_binlog_size

---

Backup

Via Velero with SeaweedFS storage:

# Milvus data is stored in SeaweedFS
# SeaweedFS is backed up via Velero to Archival S3

---

Consequences

Positive:

• Billion-scale vector search
• Hybrid dense + sparse search
• Multiple index types
• Kubernetes-native
• Active community

Negative:

• Complex distributed architecture
• Resource-intensive for large collections
• Learning curve

---

Part of OpenOva