Split View: Vector Database 완전 가이드 2025: 임베딩, 유사도 검색, Pinecone/Weaviate/Qdrant/pgvector

Traditional search: "puppy park" → keyword matching → only returns docs containing "puppy" AND "park"
Vector search:      "puppy park" → semantic vectorization → returns all semantically similar docs
                    → "a dog running on the lawn" ✅ found
                    → "best pet walking spots" ✅ found

from openai import OpenAI

client = OpenAI()

# Convert text to vector
response = client.embeddings.create(
    model="text-embedding-3-large",
    input="Vector databases are essential AI infrastructure"
)

embedding = response.data[0].embedding
print(f"Dimensions: {len(embedding)}")    # 3072
print(f"Vector sample: {embedding[:5]}")  # [0.023, -0.041, 0.017, ...]

from sentence_transformers import SentenceTransformer
from PIL import Image

model = SentenceTransformer("clip-ViT-B-32")

# Image embedding
img = Image.open("cat_photo.jpg")
img_embedding = model.encode(img)

# Text embedding (same vector space)
text_embedding = model.encode("a cute orange cat")

# Cross-modal search is possible!
from numpy import dot
from numpy.linalg import norm

similarity = dot(img_embedding, text_embedding) / (
    norm(img_embedding) * norm(text_embedding)
)
print(f"Similarity: {similarity:.4f}")  # 0.28+ (higher if related)

Text: "sunset over the ocean"  ──┐
                                  ├──→ Same vector space → similarity comparison possible
Image: (sunset photo)           ──┘

import numpy as np

def cosine_similarity(a, b):
    return np.dot(a, b) / (np.linalg.norm(a) * np.linalg.norm(b))

# Example: 3-dimensional vectors
vec_a = np.array([1, 2, 3])
vec_b = np.array([2, 4, 6])  # Same direction, different magnitude
vec_c = np.array([-1, -2, -3])  # Opposite direction

print(cosine_similarity(vec_a, vec_b))  # 1.0 (exact same direction)
print(cosine_similarity(vec_a, vec_c))  # -1.0 (exact opposite)

def euclidean_distance(a, b):
    return np.linalg.norm(a - b)

vec_a = np.array([1, 2])
vec_b = np.array([4, 6])

print(euclidean_distance(vec_a, vec_b))  # 5.0

def dot_product(a, b):
    return np.dot(a, b)

vec_a = np.array([1, 2, 3])
vec_b = np.array([4, 5, 6])

print(dot_product(vec_a, vec_b))  # 32

def manhattan_distance(a, b):
    return np.sum(np.abs(a - b))

vec_a = np.array([1, 2, 3])
vec_b = np.array([4, 6, 3])

print(manhattan_distance(vec_a, vec_b))  # 7 (3 + 4 + 0)

Text embedding search     → Cosine Similarity (default recommendation)
Image similarity search   → Euclidean Distance
Recommendation (MIPS)     → Dot Product
High-dimensional sparse   → Cosine Similarity
Clustering / Classification → Euclidean Distance

Layer 3: [A] ────────────────── [B]     (few nodes, long-range links)
           \                   /
Layer 2: [A] ── [C] ── [D] ── [B]      (medium nodes)
           \   / \   / \   / \
Layer 1: [A]-[E]-[C]-[F]-[D]-[G]-[B]   (most nodes, short-range links)
           |   |   |   |   |   |   |
Layer 0: [All vectors exist at base layer]  (all nodes)

# HNSW configuration in Qdrant
from qdrant_client import QdrantClient
from qdrant_client.models import VectorParams, HnswConfigDiff

client = QdrantClient("localhost", port=6333)

client.create_collection(
    collection_name="documents",
    vectors_config=VectorParams(
        size=1536,
        distance="Cosine"
    ),
    hnsw_config=HnswConfigDiff(
        m=16,                    # connections per node (higher = more accurate, more memory)
        ef_construct=128,        # search scope during index building
        full_scan_threshold=10000  # brute-force below this count
    )
)

Vector space:
┌─────────────────────────────────┐
│  * *   X *                      │
│   *  Cluster1  *  *             │
│  *  *          X                │
│        * *  Cluster2  *         │
│         *     *  * *            │
│              *   *              │
│    * *  *         X             │
│     Cluster3        Cluster4    │
│    *  * *         *  *  *       │
└─────────────────────────────────┘
X = cluster centroid

Search: only search within the nprobe closest clusters to the query

# Creating IVF-PQ index with FAISS
import faiss

dimension = 1536
nlist = 256      # number of clusters
m_pq = 48        # number of sub-vectors (splits dimensions into m_pq parts)
nbits = 8        # codebook size per sub-vector

quantizer = faiss.IndexFlatL2(dimension)
index = faiss.IndexIVFPQ(quantizer, dimension, nlist, m_pq, nbits)

# Train clusters + codebook on training data
index.train(training_vectors)
index.add(all_vectors)

# Set nprobe at search time
index.nprobe = 16  # number of clusters to probe
distances, indices = index.search(query_vector, k=10)

Original (float32, 1536 dims): 1536 x 4 bytes = 6,144 bytes/vector
PQ compressed (48 sub-vectors): 48 x 1 byte  = 48 bytes/vector
→ 128x compression! 100M vectors: 614GB → 4.8GB

from pinecone import Pinecone, ServerlessSpec

pc = Pinecone(api_key="YOUR_KEY")

# Create index
pc.create_index(
    name="my-index",
    dimension=1536,
    metric="cosine",
    spec=ServerlessSpec(
        cloud="aws",
        region="us-east-1"
    )
)

index = pc.Index("my-index")

# Upsert vectors
index.upsert(
    vectors=[
        ("id1", [0.1, 0.2, ...], {"title": "Document Title", "category": "tech"}),
        ("id2", [0.3, 0.4, ...], {"title": "Another Doc", "category": "science"}),
    ],
    namespace="articles"
)

# Search (filter + vector)
results = index.query(
    vector=[0.15, 0.25, ...],
    top_k=5,
    filter={"category": "tech"},
    namespace="articles",
    include_metadata=True
)

import weaviate
from weaviate.classes.config import Configure, Property, DataType

client = weaviate.connect_to_local()

# Create collection (built-in vectorization)
collection = client.collections.create(
    name="Article",
    vectorizer_config=Configure.Vectorizer.text2vec_openai(
        model="text-embedding-3-small"
    ),
    properties=[
        Property(name="title", data_type=DataType.TEXT),
        Property(name="content", data_type=DataType.TEXT),
        Property(name="category", data_type=DataType.TEXT),
    ]
)

# Add data (automatic vectorization)
articles = client.collections.get("Article")
articles.data.insert(
    properties={
        "title": "Vector DB Guide",
        "content": "Vector databases are essential for AI...",
        "category": "tech"
    }
)

# Semantic search
response = articles.query.near_text(
    query="artificial intelligence infrastructure",
    limit=5,
    filters=weaviate.classes.query.Filter.by_property("category").equal("tech")
)

from qdrant_client import QdrantClient
from qdrant_client.models import (
    Distance, VectorParams, PointStruct,
    Filter, FieldCondition, MatchValue
)

client = QdrantClient("localhost", port=6333)

# Create collection
client.create_collection(
    collection_name="articles",
    vectors_config=VectorParams(
        size=1536,
        distance=Distance.COSINE
    )
)

# Upsert vectors
client.upsert(
    collection_name="articles",
    points=[
        PointStruct(
            id=1,
            vector=[0.1, 0.2, ...],
            payload={"title": "Vector DB Guide", "category": "tech", "views": 1500}
        ),
    ]
)

# Filter + vector search
results = client.search(
    collection_name="articles",
    query_vector=[0.15, 0.25, ...],
    query_filter=Filter(
        must=[
            FieldCondition(key="category", match=MatchValue(value="tech")),
        ]
    ),
    limit=5
)

from pymilvus import connections, Collection, FieldSchema, CollectionSchema, DataType

connections.connect("default", host="localhost", port="19530")

# Define schema
fields = [
    FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
    FieldSchema(name="title", dtype=DataType.VARCHAR, max_length=200),
    FieldSchema(name="embedding", dtype=DataType.FLOAT_VECTOR, dim=1536),
]
schema = CollectionSchema(fields, description="Articles")
collection = Collection("articles", schema)

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

# Search
collection.load()
results = collection.search(
    data=[[0.1, 0.2, ...]],
    anns_field="embedding",
    param={"metric_type": "COSINE", "params": {"ef": 128}},
    limit=5,
    output_fields=["title"]
)

import chromadb

client = chromadb.PersistentClient(path="./chroma_db")

collection = client.create_collection(
    name="articles",
    metadata={"hnsw:space": "cosine"}
)

# Add documents (automatic embedding)
collection.add(
    documents=["Vector DBs are core AI infrastructure", "RAG is retrieval augmented generation"],
    metadatas=[{"category": "tech"}, {"category": "ai"}],
    ids=["doc1", "doc2"]
)

# Search
results = collection.query(
    query_texts=["artificial intelligence database"],
    n_results=5,
    where={"category": "tech"}
)

-- Requires PostgreSQL 14+
-- Install extension
CREATE EXTENSION IF NOT EXISTS vector;

-- Create table with vector column
CREATE TABLE documents (
    id BIGSERIAL PRIMARY KEY,
    title TEXT NOT NULL,
    content TEXT,
    category TEXT,
    embedding VECTOR(1536),  -- OpenAI text-embedding-3-small dimensions
    created_at TIMESTAMPTZ DEFAULT NOW()
);

-- Insert vector
INSERT INTO documents (title, content, category, embedding)
VALUES (
    'Vector DB Guide',
    'Vector databases are essential for AI...',
    'tech',
    '[0.1, 0.2, 0.3, ...]'::vector  -- 1536-dimensional vector
);

-- Cosine distance search (top 5 most similar)
SELECT id, title, 1 - (embedding <=> '[0.15, 0.25, ...]'::vector) AS similarity
FROM documents
WHERE category = 'tech'
ORDER BY embedding <=> '[0.15, 0.25, ...]'::vector
LIMIT 5;

-- L2 distance search
SELECT id, title, embedding <-> '[0.15, 0.25, ...]'::vector AS distance
FROM documents
ORDER BY embedding <-> '[0.15, 0.25, ...]'::vector
LIMIT 5;

-- Inner product search (Maximum Inner Product)
SELECT id, title, (embedding <#> '[0.15, 0.25, ...]'::vector) * -1 AS inner_product
FROM documents
ORDER BY embedding <#> '[0.15, 0.25, ...]'::vector
LIMIT 5;

-- HNSW index (recommended)
CREATE INDEX ON documents
USING hnsw (embedding vector_cosine_ops)
WITH (m = 16, ef_construction = 128);

-- Set ef_search at query time
SET hnsw.ef_search = 100;

-- IVFFlat index
CREATE INDEX ON documents
USING ivfflat (embedding vector_cosine_ops)
WITH (lists = 100);  -- number of clusters (sqrt(rows) recommended)

-- Set probes at query time
SET ivfflat.probes = 10;

-- 1. Partial index (specific category only)
CREATE INDEX ON documents
USING hnsw (embedding vector_cosine_ops)
WHERE category = 'tech';

-- 2. Filter + vector search with denormalization
-- Slow pattern: vector search then filter
SELECT * FROM documents
WHERE category = 'tech'
ORDER BY embedding <=> query_vec
LIMIT 10;

-- Fast pattern: use partitioned tables
CREATE TABLE documents_tech PARTITION OF documents
FOR VALUES IN ('tech');

-- 3. Confirm index usage with EXPLAIN
EXPLAIN (ANALYZE, BUFFERS)
SELECT id, title
FROM documents
ORDER BY embedding <=> '[0.15, 0.25, ...]'::vector
LIMIT 5;

-- Verify: Index Scan using documents_embedding_idx

import psycopg2
from pgvector.psycopg2 import register_vector

conn = psycopg2.connect("postgresql://user:pass@localhost/mydb")
register_vector(conn)

cur = conn.cursor()

# Vector search
query_embedding = [0.1, 0.2, ...]  # 1536 dimensions
cur.execute("""
    SELECT id, title, 1 - (embedding <=> %s::vector) AS similarity
    FROM documents
    WHERE category = %s
    ORDER BY embedding <=> %s::vector
    LIMIT %s
""", (query_embedding, "tech", query_embedding, 5))

results = cur.fetchall()
for row in results:
    print(f"ID: {row[0]}, Title: {row[1]}, Similarity: {row[2]:.4f}")

Query: "PostgreSQL 16 release notes"

Vector only: "MySQL 8.0 new features" may also rank high (similar semantics)
Keyword only: only documents containing exactly "PostgreSQL 16"
Hybrid: semantically similar + contains "PostgreSQL 16" = optimal results

# Hybrid search in Weaviate
response = articles.query.hybrid(
    query="PostgreSQL vector search performance",
    alpha=0.5,  # 0 = keyword only, 1 = vector only, 0.5 = balanced
    limit=10,
    fusion_type="relative_score"  # rankedFusion or relativeScoreFusion
)

def reciprocal_rank_fusion(keyword_results, vector_results, k=60):
    """
    RRF score = sum(1 / (k + rank_i))
    k = 60 is standard
    """
    scores = {}
    
    for rank, doc_id in enumerate(keyword_results):
        scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank + 1)
    
    for rank, doc_id in enumerate(vector_results):
        scores[doc_id] = scores.get(doc_id, 0) + 1 / (k + rank + 1)
    
    # Sort by score
    return sorted(scores.items(), key=lambda x: x[1], reverse=True)

# Example
keyword_hits = ["doc_A", "doc_B", "doc_C", "doc_D"]
vector_hits = ["doc_C", "doc_A", "doc_E", "doc_B"]

fused = reciprocal_rank_fusion(keyword_hits, vector_hits)
# doc_A and doc_C appear top in both → final top results

-- Hybrid search (PostgreSQL)
WITH vector_search AS (
    SELECT id, title, content,
           1 - (embedding <=> query_vec) AS vector_score,
           ROW_NUMBER() OVER (ORDER BY embedding <=> query_vec) AS vector_rank
    FROM documents
    ORDER BY embedding <=> query_vec
    LIMIT 20
),
keyword_search AS (
    SELECT id, title, content,
           ts_rank(to_tsvector('english', content), plainto_tsquery('english', 'vector database')) AS bm25_score,
           ROW_NUMBER() OVER (ORDER BY ts_rank(to_tsvector('english', content), plainto_tsquery('english', 'vector database')) DESC) AS keyword_rank
    FROM documents
    WHERE to_tsvector('english', content) @@ plainto_tsquery('english', 'vector database')
    LIMIT 20
)
SELECT
    COALESCE(v.id, k.id) AS id,
    COALESCE(v.title, k.title) AS title,
    -- RRF score
    COALESCE(1.0 / (60 + v.vector_rank), 0) +
    COALESCE(1.0 / (60 + k.keyword_rank), 0) AS rrf_score
FROM vector_search v
FULL OUTER JOIN keyword_search k ON v.id = k.id
ORDER BY rrf_score DESC
LIMIT 10;

Pre-filtering:  Apply filter → search only filtered vectors (accurate but can be slow)
Post-filtering: Vector search → apply filter to results (fast but may return fewer results)
In-filtering:   Apply filter during search simultaneously (optimal but complex to implement)

from qdrant_client.models import Filter, FieldCondition, MatchValue, Range

# Complex filter
results = client.search(
    collection_name="products",
    query_vector=query_embedding,
    query_filter=Filter(
        must=[
            FieldCondition(key="category", match=MatchValue(value="electronics")),
            FieldCondition(key="price", range=Range(gte=100, lte=500)),
        ],
        must_not=[
            FieldCondition(key="out_of_stock", match=MatchValue(value=True)),
        ],
        should=[
            FieldCondition(key="brand", match=MatchValue(value="Apple")),
            FieldCondition(key="brand", match=MatchValue(value="Samsung")),
        ]
    ),
    limit=10
)

Recommended:
- Create payload indexes on frequently filtered fields
- Works best for low-cardinality fields (category, status)
- Date range filters need indexes
- Prefer flat structures over nested objects

Not recommended:
- Indexing fields with very high cardinality (user IDs, etc.)
- Using only metadata for search without vectors (a regular DB is better)

Strategy 1: Namespace/Partition Separation (Pinecone, Milvus)
┌──────────── Index ────────────┐
│  Namespace: tenant_A  *****   │
│  Namespace: tenant_B  ooooo   │
│  Namespace: tenant_C  ^^^^^   │
└───────────────────────────────┘
- Pros: Simple, data isolation
- Cons: Cross-tenant search difficult

Strategy 2: Metadata Filter (Qdrant, Weaviate)
┌──────── Collection ───────────┐
│  *(A) o(B) ^(C) *(A) o(B)    │
│  ^(C) *(A) *(A) o(B) ^(C)    │
│  Separated by tenant_id filter│
└───────────────────────────────┘
- Pros: Flexible, cross-search possible
- Cons: Performance degrades with many tenants

Strategy 3: Collection Separation (few large tenants)
┌─ Collection: tenant_A ─┐
│  ******************     │
└────────────────────────┘
┌─ Collection: tenant_B ─┐
│  oooooooooooooooooo     │
└────────────────────────┘
- Pros: Perfect isolation, independent scaling
- Cons: Management overhead

# Qdrant example: collection management
# List collections
collections = client.get_collections()

# Collection info
info = client.get_collection("articles")
print(f"Vector count: {info.points_count}")
print(f"Index status: {info.status}")

# Delete collection
client.delete_collection("old_articles")

# Aliases (blue-green deployment)
client.update_collection_aliases(
    change_aliases_operations=[
        {"create_alias": {"collection_name": "articles_v2", "alias_name": "articles_prod"}}
    ]
)

Horizontal Sharding: distribute data across multiple nodes
┌─ Node 1 ─┐ ┌─ Node 2 ─┐ ┌─ Node 3 ─┐
│ Shard 1   │ │ Shard 2   │ │ Shard 3   │
│ 33% data  │ │ 33% data  │ │ 33% data  │
└───────────┘ └───────────┘ └───────────┘
        ↕ Query all nodes at search time, then merge

Qdrant sharding config:
- auto sharding: automatically distributes based on shard_number
- custom sharding: tenant-level separation via shard_key

# Qdrant replication setup
client.create_collection(
    collection_name="articles",
    vectors_config=VectorParams(size=1536, distance=Distance.COSINE),
    replication_factor=3,        # replicate to 3 nodes
    write_consistency_factor=2   # write succeeds after 2 nodes confirm
)

# Qdrant snapshot creation
curl -X POST "http://localhost:6333/collections/articles/snapshots"

# Snapshot recovery
curl -X PUT "http://localhost:6333/collections/articles/snapshots/recover" \
  -H "Content-Type: application/json" \
  -d '{"location": "http://backup-server/snapshot.tar"}'

# pgvector: use pg_dump
pg_dump -t documents mydb > documents_backup.sql

# Key metrics
monitoring_metrics = {
    "Search latency (p50, p95, p99)": "Target: p99 under 100ms",
    "QPS (queries per second)": "Monitor under load",
    "Recall@K": "Accuracy. Target 0.95+",
    "Index size / memory usage": "Prevent OOM",
    "Insert latency": "Critical for real-time updates",
    "Disk usage": "Capacity planning",
    "Replication lag": "Data consistency",
}