Feast + Cassandra, setup

The following example shows how to connect your prompt templates with the Feast feature store with minimal boilerplate.

Feast and Cassandra

Feast can admit several database technologies as its storage layer, one of which being Cassandra.

In this spirit, aiming at providing comprehensive information on how to integrate LangChain and Cassandra, we chose to cover this use case as well.

In order to test the code yourself, you first need to create a sample feature store. The following instructions will guide you through the setup of the very feature store (including sample data) that is used in the next code example.

Warning

The Feast example is not available as Colab and must be run locally. This means you'll need to go through a general setup, followed by LangChain-specific setup, before reading this page further.

Provision the Feature Store

You will create a new Feature Store and configure it to use the Astra DB instance you should already have as its "online store".

In practice, this amounts to Feast managing a couple of additional tables on your database.

Preliminaries

Navigate to the directory docs/frameworks/langchain/feast_store of this repo (which you should have cloned in earlier setup steps), activate the virtual environment for the LangChain examples and install this dependency:

pip install "feast[cassandra]>=0.26"

Warning

The latest version of LangChain brings an apparent conflict with Feast on package SQLAlchemy. Pending upgrades on the Feast side, our experience is that you will be fine by running the following command right after installing Feast: pip install "sqlalchemy>=2".

Do not worry about a message similar to "feast 0.33.1 requires SQLAlchemy[mypy]<2,>1, but you have sqlalchemy 2.0.20 which is incompatible.": the example notebook will run just fine.

Keep file ../../../../.env handy, as you will be shortly asked to provide the Secure Connect Bundle location and the keyspace name defined there.

Create the feature store

The following command starts an interactive creation of the feature store. Choose Astra DB, unless you want to use your on-premise Cassandra cluster:

feast init -t cassandra user_features

Provide the required information, paying attention to the keyspace name (cassio_tutorials if you went with the defaults). You can skip the optional parameters altogether.

Client ID and Client Secret

Provide the string literal token as "Client ID" and the value of ASTRA_DB_APPLICATION_TOKEN, found in your .env, as "Client Secret".

If needed, please refer to these instructions on how to generate a Token for your database.

A brand new feature store has been created in subdirectory user_features. (Note: if you give a different name to your store, adapt the following commands accordingly.)

Prepare data sources

This command creates the data sources for ingestion by Feast in the form of two *.parquet files:

python prepare_feast_data.py

Place the sources within the store, ready to be found by Feast:

mv *.parquet user_features/feature_repo/data/

Replace feature definitions

We have a ready-to-use feature definition file for this store. All you have to do is to copy it over the default one:

cp user_data_feature_definitions.py user_features/feature_repo/example_repo.py

Provision the store backend

This step will trigger Feast to create the required table in your Astra DB:

cd user_features/feature_repo/
feast apply

Materialize data to online store

Now you can have Feast transport the data into the (still empty) tables that constitute the online store:

DATE0=$(date -d "`date` - 10 years" "+%Y-%m-%dT%H:%M:%S")
DATE1=`date "+%Y-%m-%dT%H:%M:%S"`
feast materialize $DATE0 $DATE1

Ready to go

That's it. Now you can run the examples that require the Feast store.