LlamaIndex Vector_Stores Integration: Tablestore

Tablestore is a fully managed NoSQL cloud database service that enables storage of a massive amount of structured and semi-structured data.

This page shows how to use functionality related to the Tablestore vector database.

To use Tablestore, you must create an instance. Here are the creating instance instructions.

Example

pip install llama-index-vector-stores-tablestore

import os

import tablestore
from llama_index.core import MockEmbedding
from llama_index.core.schema import TextNode
from llama_index.core.vector_stores import (
    VectorStoreQuery,
    MetadataFilters,
    MetadataFilter,
    FilterCondition,
    FilterOperator,
)

from llama_index.vector_stores.tablestore import TablestoreVectorStore

# 1. create tablestore vector store
test_dimension_size = 4
store = TablestoreVectorStore(
    endpoint=os.getenv("end_point"),
    instance_name=os.getenv("instance_name"),
    access_key_id=os.getenv("access_key_id"),
    access_key_secret=os.getenv("access_key_secret"),
    vector_dimension=test_dimension_size,
    vector_metric_type=tablestore.VectorMetricType.VM_COSINE,
    # metadata mapping is used to filter non-vector fields.
    metadata_mappings=[
        tablestore.FieldSchema(
            "type",
            tablestore.FieldType.KEYWORD,
            index=True,
            enable_sort_and_agg=True,
        ),
        tablestore.FieldSchema(
            "time",
            tablestore.FieldType.LONG,
            index=True,
            enable_sort_and_agg=True,
        ),
    ],
)

# 2. create table and index
store.create_table_if_not_exist()
store.create_search_index_if_not_exist()

# 3. new a mock embedding for test
embedder = MockEmbedding(test_dimension_size)

# 4. prepare some docs
movies = [
    TextNode(
        id_="1",
        text="hello world",
        metadata={"type": "a", "time": 1995},
    ),
    TextNode(
        id_="2",
        text="a b c",
        metadata={"type": "a", "time": 1990},
    ),
    TextNode(
        id_="3",
        text="sky cloud table",
        metadata={"type": "a", "time": 2009},
    ),
    TextNode(
        id_="4",
        text="dog cat",
        metadata={"type": "a", "time": 2023},
    ),
    TextNode(
        id_="5",
        text="computer python java",
        metadata={"type": "b", "time": 2018},
    ),
    TextNode(
        id_="6",
        text="java python js nodejs",
        metadata={"type": "c", "time": 2010},
    ),
    TextNode(
        id_="7",
        text="sdk golang python",
        metadata={"type": "a", "time": 2023},
    ),
]
for movie in movies:
    movie.embedding = embedder.get_text_embedding(movie.text)

# 5. write some docs
ids = store.add(movies)
assert len(ids) == 7

# 6. delete docs
store.delete(ids[0])

# 7. query with filters
query_result = store.query(
    query=VectorStoreQuery(
        query_embedding=embedder.get_text_embedding("nature fight physical"),
        similarity_top_k=5,
        filters=MetadataFilters(
            filters=[
                MetadataFilter(
                    key="type", value="a", operator=FilterOperator.EQ
                ),
                MetadataFilter(
                    key="time", value=2020, operator=FilterOperator.LTE
                ),
            ],
            condition=FilterCondition.AND,
        ),
    ),
)
print(query_result)