tidb-vector-python

Use TiDB Vector Search with Python.

Usage

TiDB is a SQL database so that this package introduces Vector Search capability for Python ORMs:

• #SQLAlchemy
• #Peewee
• #Django

Pick one that you are familiar with to get started. If you are not using any of them, we recommend #SQLAlchemy.

We also provide a Vector Search client for simple usage:

• #TiDB Vector Client

SQLAlchemy

Install:

pip install tidb-vector sqlalchemy pymysql

Usage:

from sqlalchemy import Integer, Column
from sqlalchemy import create_engine, select
from sqlalchemy.dialects.mysql import LONGTEXT
from sqlalchemy.orm import Session, declarative_base

import tidb_vector
from tidb_vector.sqlalchemy import VectorType, VectorAdaptor

engine = create_engine("mysql+pymysql://[email protected]:4000/test")
Base = declarative_base()


# Define table schema
class Doc(Base):
    __tablename__ = "doc"
    id = Column(Integer, primary_key=True)
    embedding = Column(VectorType(dim=3))
    content = Column(LONGTEXT)


# Create empty table
Base.metadata.drop_all(engine)  # clean data from last run
Base.metadata.create_all(engine)

# Create index for L2 distance
VectorAdaptor(engine).create_vector_index(
    Doc.embedding, tidb_vector.DistanceMetric.L2, skip_existing=True
    # For cosine distance, use tidb_vector.DistanceMetric.COSINE
)

# Insert content with vectors
with Session(engine) as session:
    session.add(Doc(id=1, content="dog", embedding=[1, 2, 1]))
    session.add(Doc(id=2, content="fish", embedding=[1, 2, 4]))
    session.add(Doc(id=3, content="tree", embedding=[1, 0, 0]))
    session.commit()

# Perform Vector Search for Top K=1
with Session(engine) as session:
    results = session.execute(
        select(Doc.id, Doc.content)
        .order_by(Doc.embedding.l2_distance([1, 2, 3]))
        # For cosine distance, use Doc.embedding.cosine_distance(...)
        .limit(1)
    ).all()
    print(results)

# Perform filtered Vector Search by adding a Where Clause:
with Session(engine) as session:
    results = session.execute(
        select(Doc.id, Doc.content)
        .where(Doc.content == "dog")
        .order_by(Doc.embedding.l2_distance([1, 2, 3]))
        .limit(1)
    ).all()
    print(results)

Peewee

Install:

pip install tidb-vector peewee pymysql

Usage:

import tidb_vector
from peewee import Model, MySQLDatabase, IntegerField, TextField
from tidb_vector.peewee import VectorField, VectorAdaptor

db = MySQLDatabase(
    database="test",
    user="root",
    password="",
    host="127.0.0.1",
    port=4000,
)


# Define table schema
class Doc(Model):
    class Meta:
        database = db
        table_name = "peewee_test"

    id = IntegerField(primary_key=True)
    embedding = VectorField(3)
    content = TextField()


# Create empty table and index for L2 distance
db.drop_tables([Doc])  # clean data from last run
db.create_tables([Doc])
# For cosine distance, use tidb_vector.DistanceMetric.COSINE
VectorAdaptor(db).create_vector_index(Doc.embedding, tidb_vector.DistanceMetric.L2)

# Insert content with vectors
Doc.insert_many(
    [
        {"id": 1, "content": "dog", "embedding": [1, 2, 1]},
        {"id": 2, "content": "fish", "embedding": [1, 2, 4]},
        {"id": 3, "content": "tree", "embedding": [1, 0, 0]},
    ]
).execute()

# Perform Vector Search for Top K=1
cursor = (
    Doc.select(Doc.id, Doc.content)
    # For cosine distance, use Doc.embedding.cosine_distance(...)
    .order_by(Doc.embedding.l2_distance([1, 2, 3]))
    .limit(1)
)
for row in cursor:
    print(row.id, row.content)


# Perform filtered Vector Search by adding a Where Clause:
cursor = (
    Doc.select(Doc.id, Doc.content)
    .where(Doc.content == "dog")
    .order_by(Doc.embedding.l2_distance([1, 2, 3]))
    .limit(1)
)
for row in cursor:
    print(row.id, row.content)

Django

Tip

Django is a full-featured web framework, not just an ORM. The following usage introducutions are provided for existing Django users.

For new users to get started, consider using SQLAlchemy or Peewee.

Install:

pip install 'django-tidb[vector]~=5.0.0' 'django~=5.0.0'  mysqlclient

Usage:

1. Configure django_tidb as engine, like:

DATABASES = {
    'default': {
        'ENGINE': 'django_tidb',
        'NAME': 'django',
        'USER': 'root',
        'PASSWORD': '',
        'HOST': '127.0.0.1',
        'PORT': 4000,
    },
}

2. Define a model with a vector field and vector index:

from django.db import models
from django_tidb.fields.vector import VectorField, VectorIndex, L2Distance

class Doc(models.Model):
    id = models.IntegerField(primary_key=True)
    embedding = VectorField(dimensions=3)
    content = models.TextField()
    class Meta:
        indexes = [VectorIndex(L2Distance("embedding"), name="idx")]

3. Insert data:

Doc.objects.create(id=1, content="dog", embedding=[1, 2, 1])
Doc.objects.create(id=2, content="fish", embedding=[1, 2, 4])
Doc.objects.create(id=3, content="tree", embedding=[1, 0, 0])

4. Perform Vector Search for Top K=1:

queryset = (
    Doc.objects
        .order_by(L2Distance("embedding", [1, 2, 3]))
        .values("id", "content")[:1]
)
print(queryset)

5. Perform filtered Vector Search by adding a Where Clause:

queryset = (
     Doc.objects
          .filter(content="dog")
          .order_by(L2Distance("embedding", [1, 2, 3]))
          .values("id", "content")[:1]
)
print(queryset)

For more details, see django-tidb.

TiDB Vector Client

Within the framework, you can directly utilize the built-in TiDBVectorClient, as demonstrated by integrations like Langchain and Llama index, to seamlessly interact with TiDB Vector. This approach abstracts away the need to manage the underlying ORM, simplifying your interaction with the vector store.

We provide TiDBVectorClient which is based on sqlalchemy, you need to use pip install tidb-vector[client] to install it.

Create a TiDBVectorClient instance:

from tidb_vector.integrations import TiDBVectorClient

TABLE_NAME = 'vector_test'
CONNECTION_STRING = 'mysql+pymysql://<USER>:<PASSWORD>@<HOST>:4000/<DB>?ssl_verify_cert=true&ssl_verify_identity=true'

tidb_vs = TiDBVectorClient(
    # the table which will store the vector data
    table_name=TABLE_NAME,
    # tidb connection string
    connection_string=CONNECTION_STRING,
    # the dimension of the vector, in this example, we use the ada model, which has 1536 dimensions
    vector_dimension=1536,
    # if recreate the table if it already exists
    drop_existing_table=True,
)

Bulk insert:

ids = [
    "f8e7dee2-63b6-42f1-8b60-2d46710c1971",
    "8dde1fbc-2522-4ca2-aedf-5dcb2966d1c6",
    "e4991349-d00b-485c-a481-f61695f2b5ae",
]
documents = ["foo", "bar", "baz"]
embeddings = [
    text_to_embedding("foo"),
    text_to_embedding("bar"),
    text_to_embedding("baz"),
]
metadatas = [
    {"page": 1, "category": "P1"},
    {"page": 2, "category": "P1"},
    {"page": 3, "category": "P2"},
]

tidb_vs.insert(
    ids=ids,
    texts=documents,
    embeddings=embeddings,
    metadatas=metadatas,
)

Query:

tidb_vs.query(text_to_embedding("foo"), k=3)

# query with filter
tidb_vs.query(text_to_embedding("foo"), k=3, filter={"category": "P1"})

Bulk delete:

tidb_vs.delete(["f8e7dee2-63b6-42f1-8b60-2d46710c1971"])

# delete with filter
tidb_vs.delete(["f8e7dee2-63b6-42f1-8b60-2d46710c1971"], filter={"category": "P1"})

Examples

There are some examples to show how to use the tidb-vector-python to interact with TiDB Vector in different scenarios.

• OpenAI Embedding: use the OpenAI embedding model to generate vectors for text data, store them in TiDB Vector, and search for similar text.
• Image Search: use the OpenAI CLIP model to generate vectors for image and text, store them in TiDB Vector, and search for similar images.
• LlamaIndex RAG with UI: use the LlamaIndex to build an RAG(Retrieval-Augmented Generation) application.
• Chat with URL: use LlamaIndex to build an RAG(Retrieval-Augmented Generation) application that can chat with a URL.
• GraphRAG: 20 lines code of using TiDB Serverless to build a Knowledge Graph based RAG application.
• GraphRAG Step by Step Tutorial: Step by step tutorial to build a Knowledge Graph based RAG application with Colab notebook. In this tutorial, you will learn how to extract knowledge from a text corpus, build a Knowledge Graph, store the Knowledge Graph in TiDB Serverless, and search from the Knowledge Graph.
• Vector Search Notebook with SQLAlchemy: use SQLAlchemy to interact with TiDB Serverless: connect db, index&store data and then search vectors.
• Build RAG with Jina AI Embeddings: use Jina AI to generate embeddings for text data, store the embeddings in TiDB Vector Storage, and search for similar embeddings.
• Semantic Cache: build a semantic cache with Jina AI and TiDB Vector.

for more examples, see the examples directory.

Contributing

Please feel free to reach out to the maintainers if you have any questions or need help with the project. Before contributing, please read the CONTRIBUTING.md file.