SQLite向量搜索：最轻量的RAG方案

向量搜索真的需要服务器吗？sqlite-vec把向量加到最便携的数据库里。

SQLite Vector Search 2026: sqlite-vec and the Lightest RAG Stack Ever

SQLite is the most deployed database on Earth - every phone, browser, and desktop app ships with it. sqlite-vec (7,900 GitHub stars) adds vector search as a loadable extension, making it possible to run a full RAG pipeline inside a single file. No server, no Docker, no cloud.

Why SQLite for Vectors

• Zero infrastructure - one file, no daemon
• Perfect for desktop/mobile apps - embed RAG in your app
• Battery-friendly - no network calls for local semantic search
• Good to ~1M vectors - plenty for personal or small-team use

Step 1: Install

pip install sqlite-vec
# or load as a SQLite extension in any language

Step 2: Create and Insert

import sqlite3, sqlite_vec

conn = sqlite3.connect("search.db")
conn.enable_load_extension(True)
sqlite_vec.load(conn)

conn.execute("CREATE VIRTUAL TABLE vec_docs USING vec0(embedding float[768])")

# Insert with an embedding from any model
conn.execute(
    "INSERT INTO vec_docs(rowid, embedding) VALUES (?, ?)",
    (1, embedding_list)
)

Step 3: Search

rows = conn.execute(
    "SELECT rowid, distance FROM vec_docs WHERE embedding MATCH ? ORDER BY distance LIMIT 5",
    (query_embedding,)
).fetchall()

Step 4: Full RAG in One File

The complete stack: documents in a normal table, embeddings in the vec0 virtual table, and the LLM served locally by Ollama. Total: one .db file + one Python script.

When to Use It vs a Real Vector DB

FAQ

sqlite-vec vs sqlite-vss? sqlite-vss is the older, unmaintained project; sqlite-vec is its modern successor with better performance and maintenance.

Does it work on mobile? Yes - loads in iOS and Android apps; it is written in C with no server component.

How many vectors can it handle? Community benchmarks show smooth operation to ~1M vectors; beyond that, a server-based vector DB is a better fit.