PGVectorStore

PGVectorStore is an implementation of a LangChain vectorstore using postgres as the backend. This notebook goes over how to use the PGVectorStore API. The code lives in an integration package called: langchain-postgres.

Setup

This package requires a PostgreSQL database with the pgvector extension. You can run the following command to spin up a container for a pgvector enabled Postgres instance:

docker run --name pgvector-container -e POSTGRES_USER=langchain -e POSTGRES_PASSWORD=langchain -e POSTGRES_DB=langchain -p 6024:5432 -d pgvector/pgvector:pg16

Install

Install the integration library, langchain-postgres.

%pip install -qU  langchain-postgres
# This notebook also requires the following dependencies
%pip install -qU  langchain-core langchain-cohere sqlalchemy

Set your Postgres values

Set your Postgres values to test the functionality in this notebook against a Postgres instance.

# @title Set your values or use the defaults to connect to Docker { display-mode: "form" }
POSTGRES_USER = "langchain"  # @param {type: "string"}
POSTGRES_PASSWORD = "langchain"  # @param {type: "string"}
POSTGRES_HOST = "localhost"  # @param {type: "string"}
POSTGRES_PORT = "6024"  # @param {type: "string"}
POSTGRES_DB = "langchain"  # @param {type: "string"}
TABLE_NAME = "vectorstore"  # @param {type: "string"}
VECTOR_SIZE = 1024  # @param {type: "int"}

Initialization

PGEngine Connection Pool

One of the requirements and arguments to establish PostgreSQL as a vector store is a PGEngine object. The PGEngine configures a shared connection pool to your Postgres database. This is an industry best practice to manage number of connections and to reduce latency through cached database connections. PGVectorStore can be used with the asyncpg and psycopg3 drivers. To create a PGEngine using PGEngine.from_connection_string() you need to provide:

url : Connection string using the postgresql+asyncpg driver.

Note: This tutorial demonstrates the async interface. All async methods have corresponding sync methods.

# See docker command above to launch a Postgres instance with pgvector enabled.
CONNECTION_STRING = (
    f"postgresql+asyncpg://{POSTGRES_USER}:{POSTGRES_PASSWORD}@{POSTGRES_HOST}"
    f":{POSTGRES_PORT}/{POSTGRES_DB}"
)
# To use psycopg3 driver, set your connection string to `postgresql+psycopg://`

from langchain_postgres import PGEngine

pg_engine = PGEngine.from_connection_string(url=CONNECTION_STRING)

To create a PGEngine using PGEngine.from_engine() you need to provide:

engine : An object of AsyncEngine

from sqlalchemy.ext.asyncio import create_async_engine

# Create an SQLAlchemy Async Engine
engine = create_async_engine(
    CONNECTION_STRING,
)

pg_engine = PGEngine.from_engine(engine=engine)

Initialize a table

The PGVectorStore class requires a database table. The PGEngine engine has a helper method ainit_vectorstore_table() that can be used to create a table with the proper schema for you. See Create a custom Vector Store or Create a Vector Store using existing table for customizing the schema.

await pg_engine.ainit_vectorstore_table(
    table_name=TABLE_NAME,
    vector_size=VECTOR_SIZE,
)

Optional Tip: 💡

You can also specify a schema name by passing schema_name wherever you pass table_name. Eg:

SCHEMA_NAME="my_schema"

await pg_engine.ainit_vectorstore_table(
    table_name=TABLE_NAME,
    vector_size=768,
    schema_name=SCHEMA_NAME,    # Default: "public"
)

Create an embedding class instance

You can use any LangChain embeddings model.

from langchain_cohere import CohereEmbeddings

embedding = CohereEmbeddings(model="embed-english-v3.0")

Initialize a default PGVectorStore

Use the default table schema to connect to the vectorstore. See Create a custom Vector Store or Create a Vector Store using existing table for customizing the schema.

from langchain_postgres import PGVectorStore

store = await PGVectorStore.create(
    engine=pg_engine,
    table_name=TABLE_NAME,
    # schema_name=SCHEMA_NAME,
    embedding_service=embedding,
)

Manage vector store

Add documents

Add documents to the vector store. Metadata is stored in a JSON column, see “Create a custom Vector Store” to store metadata to be used for filters.

import uuid

from langchain_core.documents import Document

docs = [
    Document(
        id=str(uuid.uuid4()),
        page_content="Red Apple",
        metadata={"description": "red", "content": "1", "category": "fruit"},
    ),
    Document(
        id=str(uuid.uuid4()),
        page_content="Banana Cavendish",
        metadata={"description": "yellow", "content": "2", "category": "fruit"},
    ),
    Document(
        id=str(uuid.uuid4()),
        page_content="Orange Navel",
        metadata={"description": "orange", "content": "3", "category": "fruit"},
    ),
]

await store.aadd_documents(docs)

Add texts

Add text directly to the vectorstore, if not structured as a Document.

import uuid

all_texts = ["Apples and oranges", "Cars and airplanes", "Pineapple", "Train", "Banana"]
metadatas = [{"len": len(t)} for t in all_texts]
ids = [str(uuid.uuid4()) for _ in all_texts]

await store.aadd_texts(all_texts, metadatas=metadatas, ids=ids)

Delete documents

Documents can be deleted using ids.

await store.adelete([ids[1]])

Query vector store

Search for documents

Use a natural language query to search for similar documents.

query = "I'd like a fruit."
docs = await store.asimilarity_search(query)
print(docs)

Search for documents by vector

Search for similar documents using a vector embedding.

query_vector = embedding.embed_query(query)
docs = await store.asimilarity_search_by_vector(query_vector, k=2)
print(docs)

Add a Index

Speed up vector search queries by applying a vector index. Learn more about vector indexes. Indexes will use a default index name if a name is not provided. To add multiple indexes, different index names are required.

from langchain_postgres.v2.indexes import HNSWIndex, IVFFlatIndex

index = IVFFlatIndex()
await store.aapply_vector_index(index)

index = HNSWIndex(name="my-hnsw-index")
await store.aapply_vector_index(index)

Set index parameters to tune the index for optimal balance between recall and QPS.

index = IVFFlatIndex(name="my-ivfflat", lists=120)
await store.aapply_vector_index(index)

Re-index

Rebuild an index using the data stored in the index’s table, replacing the old copy of the index. Some index types may require re-indexing after a considerable amount of new data is added.

await store.areindex()  # Re-index using the default index name
await store.areindex("my-hnsw-index")  # Re-index using the index name

Drop an index

Remove a vector index.

await store.adrop_vector_index()  # Delete index using the default name
await store.adrop_vector_index("my-hnsw-index")  # Delete index using the index name

Create a custom Vector Store

Customize the vectorstore with special column names or with custom metadata columns. ainit_vectorstore_table

Use fields content_column, embedding_column,metadata_columns, metadata_json_column, id_column to rename the columns.
Use the Column class to create custom id or metadata columns. A Column is defined by a name and data type. Any Postgres data type can be used.
Use store_metadata to create a JSON column to store extra metadata.

Optional Tip: 💡

To use non-uuid ids, you must customize the id column:

await pg_engine.ainit_vectorstore_table(
    ...,
    id_column=Column(name="langchain_id", data_type="INTEGER")
)

PGVectorStore

Use fields content_column, embedding_column,metadata_columns, metadata_json_column, id_column to rename the columns.
ignore_metadata_columns to ignore columns that should not be used for Document metadata. This is helpful when using a preexisting table, where all data columns are not necessary.
Use a different distance_strategy for the similarity calculation during vector search.
Use index_query_options to tune local index parameters during vector search.

from langchain_postgres import Column

# Set table name
TABLE_NAME = "vectorstore_custom"
# SCHEMA_NAME = "my_schema"

await pg_engine.ainit_vectorstore_table(
    table_name=TABLE_NAME,
    # schema_name=SCHEMA_NAME,
    vector_size=VECTOR_SIZE,
    metadata_columns=[Column("len", "INTEGER")],
)


# Initialize PGVectorStore
custom_store = await PGVectorStore.create(
    engine=pg_engine,
    table_name=TABLE_NAME,
    # schema_name=SCHEMA_NAME,
    embedding_service=embedding,
    metadata_columns=["len"],
)

Search for documents with metadata filter

A Vector Store can take advantage of relational data to filter similarity searches. The vectorstore supports a set of filters that can be applied against the metadata fields of the documents. See the migration guide for details on how to migrate to use metadata columns. PGVectorStore currently supports the following operators and all Postgres data types.

Operator	Meaning/Category
$eq	Equality (==)
$ne	Inequality (!=)
$lt	Less than (<)
$lte	Less than or equal (<=)
$gt	Greater than (>)
$gte	Greater than or equal (>=)
$in	Special Cased (in)
$nin	Special Cased (not in)
$between	Special Cased (between)
$exists	Special Cased (is null)
$like	Text (like)
$ilike	Text (case-insensitive like)
$and	Logical (and)
$or	Logical (or)

import uuid

docs = [
    Document(
        id=str(uuid.uuid4()),
        page_content="Red Apple",
        metadata={"description": "red", "content": "1", "category": "fruit"},
    ),
    Document(
        id=str(uuid.uuid4()),
        page_content="Banana Cavendish",
        metadata={"description": "yellow", "content": "2", "category": "fruit"},
    ),
    Document(
        id=str(uuid.uuid4()),
        page_content="Orange Navel",
        metadata={"description": "orange", "content": "3", "category": "fruit"},
    ),
]

await custom_store.aadd_documents(docs)

# Use a dictionary filter on search
docs = await custom_store.asimilarity_search(query, filter={"content": {"$gte": 1}})

print(docs)

Create a Vector Store using existing table

A Vector Store can be built up on an existing table. Assuming there’s a pre-existing table in PG DB: products, which stores product details for an eComm venture. Here is how this table mapped to PGVectorStore:

id_column="product_id": ID column uniquely identifies each row in the products table.
content_column="description": The description column contains text descriptions of each product. This text is used by the embedding_service to create vectors that go in embedding_column and represent the semantic meaning of each description.
embedding_column="embed": The embed column stores the vectors created from the product descriptions. These vectors are used to find products with similar descriptions.
metadata_columns=["name", "category", "price_usd", "quantity", "sku", "image_url"]: These columns are treated as metadata for each product. Metadata provides additional information about a product, such as its name, category, price, quantity available, SKU (Stock Keeping Unit), and an image URL. This information is useful for displaying product details in search results or for filtering and categorization.
metadata_json_column="metadata": The metadata column can store any additional information about the products in a flexible JSON format. This allows for storing varied and complex data that doesn’t fit into the standard columns.

# Set an existing table name
TABLE_NAME = "products"
# SCHEMA_NAME = "my_schema"

# Initialize PGVectorStore
custom_store = await PGVectorStore.create(
    engine=pg_engine,
    table_name=TABLE_NAME,
    # schema_name=SCHEMA_NAME,
    embedding_service=embedding,
    # Connect to existing VectorStore by customizing below column names
    id_column="product_id",
    content_column="description",
    embedding_column="embed",
    metadata_columns=["name", "category", "price_usd", "quantity", "sku", "image_url"],
    metadata_json_column="metadata",
)

Note:

Optional: If the embed column is newly created or has different dimensions than supported by embedding model, it is required to one-time add the embeddings for the old records, like this: ALTER TABLE products ADD COLUMN embed vector(768) DEFAULT NULL
For new records, added via VectorStore embeddings are automatically generated.

Clean up

⚠️ WARNING: this can not be undone Drop the vector store table.

await pg_engine.adrop_table(TABLE_NAME)

Usage for retrieval-augmented generation

For guides on how to use this vector store for retrieval-augmented generation (RAG), see the following sections:

API reference

For detailed documentation of all VectorStore features and configurations head to the API reference: python.langchain.com/api_reference/postgres/v2/langchain_postgres.v2.vectorstores.PGVectorStore.html

Providers

Integrations by component

Setup

Install

Set your Postgres values

Initialization

PGEngine Connection Pool

Initialize a table

Optional Tip: 💡

Create an embedding class instance

Initialize a default PGVectorStore

Manage vector store

Add documents

Add texts

Delete documents

Query vector store

Search for documents

Search for documents by vector

Add a Index

Re-index

Drop an index

Create a custom Vector Store

Optional Tip: 💡

Search for documents with metadata filter

Create a Vector Store using existing table

Clean up

Usage for retrieval-augmented generation

API reference

Providers

Integrations by component

​Setup

​Install

​Set your Postgres values

​Initialization

​PGEngine Connection Pool

​Initialize a table

​Optional Tip: 💡

​Create an embedding class instance

​Initialize a default PGVectorStore

​Manage vector store

​Add documents

​Add texts

​Delete documents

​Query vector store

​Search for documents

​Search for documents by vector

​Add a Index

​Re-index

​Drop an index

​Create a custom Vector Store

​Optional Tip: 💡

​Search for documents with metadata filter

​Create a Vector Store using existing table

​Clean up

​Usage for retrieval-augmented generation

​API reference

Setup

Install

Set your Postgres values

Initialization

PGEngine Connection Pool

Initialize a table

Optional Tip: 💡

Create an embedding class instance

Initialize a default PGVectorStore

Manage vector store

Add documents

Add texts

Delete documents

Query vector store

Search for documents

Search for documents by vector

Add a Index

Re-index

Drop an index

Create a custom Vector Store

Optional Tip: 💡

Search for documents with metadata filter

Create a Vector Store using existing table

Clean up

Usage for retrieval-augmented generation

API reference