Python SDK for TiDB AI: A unified data platform empowering developers to build next-generation AI applications.

Introduction

Python SDK for TiDB AI: A unified data platform empowering developers to build next-generation AI applications.

• 🔍 Unified Search Modes: Vector · Full‑Text · Hybrid
• 🎭 Auto‑Embedding & Multi‑Modal Storage: Support for text, images, and more
• 🖼️ Image Search Support: Text‑to‑image and image‑to‑image retrieval capabilities
• 🎯 Advanced Filtering & Reranking: Flexible filters with optional reranker models to fine-tune result relevance
• 💱 Transaction Support: Full transaction management including commit/rollback to ensure consistency

Installation

[!NOTE] This Python package is under rapid development and its API may change. It is recommended to use a fixed version when installing, e.g., pytidb==0.0.9.

pip install pytidb

# To use built-in embedding functions and rerankers:
pip install "pytidb[models]"

# To convert query results to pandas DataFrame:
pip install pandas

Connect to TiDB Cloud

Create a free TiDB cluster at tidbcloud.com.

import os
from pytidb import TiDBClient

db = TiDBClient.connect(
    host=os.getenv("TIDB_HOST"),
    port=int(os.getenv("TIDB_PORT")),
    username=os.getenv("TIDB_USERNAME"),
    password=os.getenv("TIDB_PASSWORD"),
    database=os.getenv("TIDB_DATABASE"),
    ensure_db=True,
)

Highlights# 🤖 Automatic Embedding

PyTiDB automatically embeds text fields (e.g., text) and stores the vector embedding in a vector field (e.g., text_vec).

Create a table with an embedding function:

from pytidb.schema import TableModel, Field, FullTextField
from pytidb.embeddings import EmbeddingFunction

text_embed = EmbeddingFunction("openai/text-embedding-3-small")

class Chunk(TableModel):
    __tablename__ = "chunks"

    id: int = Field(primary_key=True)
    text: str = FullTextField()
    text_vec: list[float] = text_embed.VectorField(
        source_field="text"
    )  # 👈 Defines the vector field.
    user_id: int = Field()

table = db.create_table(schema=Chunk, if_exists="skip")

Bulk insert data:

table.bulk_insert([
    Chunk(id=2, text="bar", user_id=2),   # 👈 The text field is embedded and saved to text_vec automatically.
    Chunk(id=3, text="baz", user_id=3),
    Chunk(id=4, text="qux", user_id=4),
])

🔍 Search

Vector Search

Vector search finds the most relevant records based on semantic similarity, so you don't need to include all keywords explicitly in your query.

df = (
  table.search("<query>")  # 👈 The query is embedded automatically.
    .filter({"user_id": 2})
    .limit(2)
    .to_list()
)\n\n# Output: A list of dicts.

See the Vector Search example for more details.

Full-text Search

Full-text search tokenizes the query and finds the most relevant records by matching exact keywords.

df = (
  table.search("<query>", search_type="fulltext")
    .limit(2)
    .to_pydantic()
)\n\n# Output: A list of pydantic model instances.

See the Full-text Search example for more details.

Hybrid Search

Hybrid search combines exact matching from full-text search with semantic understanding from vector search, delivering more relevant and reliable results.

df = (
  table.search("<query>", search_type="hybrid")
    .limit(2)
    .to_pandas()
)\n\n# Output: A pandas DataFrame.

See the Hybrid Search example for more details.

Image Search

Image search lets you find visually similar images using natural language descriptions or another image as a reference.

from PIL import Image
from pytidb.schema import TableModel, Field
from pytidb.embeddings import EmbeddingFunction

# Define a multi-modal embedding model.
jina_embed_fn = EmbeddingFunction("jina_ai/jina-embeddings-v4")  # Using multi-modal embedding model.

class Pet(TableModel):
    __tablename__ = "pets"
    id: int = Field(primary_key=True)
    image_uri: str = Field()
    image_vec: list[float] = jina_embed_fn.VectorField(
        source_field="image_uri",
        source_type="image"
    )

table = db.create_table(schema=Pet, if_exists="skip")

# Insert sample images ...
table.insert(Pet(image_uri="path/to/shiba_inu_14.jpg"))

# Search for images using natural language
results = table.search("shiba inu dog").limit(1).to_list()

# Search for images using an image ...
query_image = Image.open("shiba_inu_15.jpg")
results = table.search(query_image).limit(1).to_pydantic()

See the Image Search example for more details.

Advanced Filtering

PyTiDB supports a variety of operators for flexible filtering:

⛓ Join Structured and Unstructured Data

from pytidb import Session
from pytidb.sql import select

# Create a table to store user data:
class User(TableModel):
    __tablename__ = "users"
    id: int = Field(primary_key=True)
    name: str = Field(max_length=20)

with Session(engine) as session:
    query = (
        select(Chunk).join(User, Chunk.user_id == User.id).where(User.name == "Alice")
    )
    chunks = session.exec(query).all()

[(c.id, c.text, c.user_id) for c in chunks]

💱 Transaction Support

PyTiDB supports transaction management, helping you avoid race conditions and ensure data consistency.

with db.session() as session:
    initial_total_balance = db.query("SELECT SUM(balance) FROM players").scalar()

    # Transfer 10 coins from player 1 to player 2
    db.execute("UPDATE players SET balance = balance - 10 WHERE id = 1")
    db.execute("UPDATE players SET balance = balance + 10 WHERE id = 2")

    session.commit()
    # or session.rollback()

    final_total_balance = db.query("SELECT SUM(balance) FROM players").scalar()
    assert final_total_balance == initial_total_balance

Extensions

• 🔌 Built-in MCP support

[!TIP] Click the button below to install TiDB MCP Server in Cursor. Then, confirm by clicking Install when prompted.