turbovec vs Pinecone：为什么自托管向量搜索能节省 90% 成本？

import turbovec

# 创建 turbovec 索引，自动使用 TurboQuant 算法
index = turbovec.Index(
    dimension=768,           # 向量维度
    metric="cosine",         # 余弦相似度
    quantization="turboquant" # 自动选择最优量化策略
)

# 添加 1000 万向量，内存占用仅 4GB
index.add(vectors)  # 10,000,000 × 768 float32 → 压缩后仅 4GB
print(f"存储压缩率: {1.0 - 4.0 / (10_000_000 * 768 * 4 / 1_073_741_824):.1%}")
# 输出: 存储压缩率: 87.5%

import turbovec
import duckdb

# 创建 DuckDB 数据库存储结构化元数据
db = duckdb.connect("knowledge_base.db")
db.execute("""
    CREATE TABLE documents (
        id INTEGER PRIMARY KEY,
        title VARCHAR,
        content TEXT,
        category VARCHAR,
        created_at TIMESTAMP,
        tags VARCHAR[]
    )
""")

# 创建 turbovec 向量索引
index = turbovec.Index(dimension=768)

# 混合查询：先关键词过滤，再向量检索
def hybrid_search(query_text, query_embedding, top_k=10):
    # 1. 关键词过滤（DuckDB）
    filtered_docs = db.execute(f"""
        SELECT id, title, content
        FROM documents
        WHERE content ILIKE '%{query_text}%'
          AND category = 'technical'
        LIMIT 100
    """).fetchall()
    
    # 2. 向量检索（turbovec）
    results = index.search(query_embedding, k=top_k)
    
    # 3. 合并结果
    return merge_results(filtered_docs, results)

from pinecone import Pinecone
import turbovec

pc = Pinecone(api_key="your_api_key")
index = pc.Index("courses-index")

# 导出所有向量
all_vectors = index.fetch(list(range(5_000_000))).vectors
embeddings = [v.values for v in all_vectors]

# 创建 turbovec 索引
tv_index = turbovec.Index(dimension=1536)
tv_index.add(embeddings)
tv_index.save("courses_tv.index")

import time

# Pinecone 查询
start = time.time()
pinecone_result = index.query(
    vector=query_embedding,
    top_k=10,
    include_metadata=True
)
pinecone_time = time.time() - start

# turbovec 查询
start = time.time()
tv_result = tv_index.search(query_embedding, k=10)
tv_time = time.time() - start

print(f"Pinecone: {pinecone_time*1000:.2f}ms")
print(f"turbovec: {tv_time*1000:.2f}ms")
print(f"速度提升: {pinecone_time/tv_time:.1f}x")

pip install turboquant duckdb

import turbovec
import numpy as np

# 初始化索引（1536 维向量，用于 OpenAI embeddings）
index = turbovec.Index(
    dimension=1536,
    metric="cosine",
    quantization="turboquant"
)

# 添加向量
vectors = np.random.rand(1_000_000, 1536).astype(np.float32)
index.add(vectors)

# 保存索引
index.save("my_index.index")

import turbovec

# 加载索引
index = turbovec.Index.load("my_index.index")

# 执行查询
query = np.random.rand(1536).astype(np.float32)
results = index.search(query, k=10)

for i, (vector, distance) in enumerate(results):
    print(f"Top {i+1}: distance={distance:.4f}")

import turbovec
import duckdb
import numpy as np

# 创建 DuckDB 数据库
db = duckdb.connect("app.db")
db.execute("""
    CREATE TABLE documents (
        id INTEGER PRIMARY KEY,
        content TEXT,
        embedding BLOB,
        metadata JSON
    )
""")

# 创建 turbovec 索引
tv_index = turbovec.Index(dimension=1536)

# 批量插入并创建索引
for doc_id, content, embedding in documents:
    db.execute(
        "INSERT INTO documents VALUES (?, ?, ?, ?)",
        [doc_id, content, embedding.tobytes(), '{"category": "tech"}']
    )
    tv_index.add(embedding)

# 混合查询
query_embedding = np.random.rand(1536).astype(np.float32)
tv_results = tv_index.search(query_embedding, k=10)
duck_results = db.execute(
    "SELECT id, content FROM documents WHERE metadata->>'category' = 'tech'"
).fetchall()