LangChain MatrixOne

This package contains the LangChain integration for MatrixOne.

Installation

You can install this package using pip or uv:

pip install langchain-matrixone
# or
uv pip install langchain-matrixone

Prerequisites

• A running MatrixOne instance that you can reach over MySQL protocol.
• Network credentials (host, port, user, password, database) with permission to create tables.
• An embedding model supported by LangChain (e.g., OpenAIEmbeddings, HuggingFaceEmbeddings, or a custom Embeddings implementation).
• The MatrixOne Python SDK (installed automatically with this package) must be able to connect with vector operations enabled.

Development

This project uses uv for dependency management.

# Install dependencies
uv sync

# Run tests
uv run pytest

Usage

from langchain_matrixone import MatrixOneVectorStore
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain_core.documents import Document

connection_args = {
    "host": "127.0.0.1",
    "port": 6001,
    "user": "root",
    "password": "111",
    "database": "langchain_demo",
}

embedder = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
vector_store = MatrixOneVectorStore(
    embedding=embedder,
    connection_args=connection_args,
    table_name="langchain_vectors",
)

# Ensure you have a running MatrixOne deployment accessible via MySQL.
splitter = RecursiveCharacterTextSplitter(chunk_size=400, chunk_overlap=20)
docs = splitter.split_documents(
    [
        Document(page_content="MatrixOne is a scalable cloud-native database."),
        Document(page_content="LangChain provides common interfaces for LLM apps."),
    ]
)

vector_store.add_texts([doc.page_content for doc in docs], metadatas=[doc.metadata for doc in docs])

query = "What is LangChain?"
similar = vector_store.similarity_search(query, k=2)
for doc in similar:
    print(doc.page_content, doc.metadata)

Using an existing MatrixOne Client

If you already manage a matrixone.Client elsewhere in your application, pass it directly to the vector store to reuse pooled connections:

from matrixone import Client

client = Client()
client.connect(**connection_args)

vector_store = MatrixOneVectorStore(
    embedding=embedder,
    client=client,
    table_name="langchain_vectors",
)

Notes

• The vector table schema is created automatically if it does not exist. Set drop_old=True during initialization to recreate it.
• Any additional metadata stored alongside texts must be JSON serializable.
• MatrixOne currently expects embeddings as VECF32. Ensure the embedding dimension stays constant for a given table.
• Use MatrixOneVectorStore.from_texts when you want a single call that both creates the store and inserts documents.
• For advanced vector indexing (IVF/HNSW) or additional capabilities review the MatrixOne vector operations guide.

Features

• Vector Search with MatrixOne
• Metadata filtering (if supported)
• Add/Delete documents