feast-dev
feat: Add feast rag retriver functionality
What this PR does / why we need it:
This PR adds a FeastRagRetriever that inherits from the HuggingFace Transformers RagRetriever class. It allows for integration with feast functionality allowing RAG to be performed more easily using the Feature Store.
The following has been added:
rag_retriever.py This module implements a custom RAG retriever by combining Feast feature store capabilities with transformer-based models. The implementation provides:
- A custom RAG retriever (FeastRAGRetriever) that supports three search modes:
- Text-based search
- Vector-based search
- Hybrid search
- Integration with Hugging Face's transformers library and sentence-transformers
- Configurable document formatting options
vector_store.py This module implements a FeastVectorStore class. The FeastVectorStore takes in a FeatureStore, FeatureView and features using these to query the store via the existing retrieve_online_documents_v2 function.
setup.py
Adding entries for RAG dependencies allowing for installation with pip install feast[rag].
init.py Adding entries for new classes added in rag_retriever.py and vector_store.py
pyproject.toml The dependencies required for the RAG implementation have been added here. The relevant dependencies have also been added to the relevant requirements.txt files.
feature_store.py Changes were made to the def _validate_vector_features function. The updated function now correctly validates the length of each individual vector within a specified DataFrame column, rather than checking the total number of DataFrame columns. This change ensures that every single vector matches its expected dimension, significantly improving data integrity and error reporting. This was improved following addition of unit tests.
Unit tests Unit tests have been included to cover the added functionality, along with some minor changes in the existing example_feature_repo_1.py file.
examples/rag-retriever An example has been added here including, feature_repo, README, low level design image and example notebook. This currently includes implementation on Red Hat OpenShift with a remote Milvus instance.
.github/workflows The change here is a small update suggested by @ntkathole for torch installation and error handling.
Which issue(s) this PR fixes:
Related to https://github.com/feast-dev/feast/issues/5391 but get_top_docs has not been implemented here as the FeastIndex is a dummy index - the retrieval functionality exists in the FeastRagRetriever via the FeastVector store which uses retrieve_online_documents_v2 .
Misc
LGTM. Please address @ntkathole comments and fix unit tests!
Thanks @jyejare I have addressed all comments now. The unit tests are still failing
FAILED sdk/python/tests/unit/infra/online_store/test_dynamodb_online_store.py::test_dynamodb_online_store_update - botocore.exceptions.ClientError: An error occurred (UnrecognizedClientException) when calling the ListTagsOfResource operation: The security token included in the request is invalid.
FAILED sdk/python/tests/unit/infra/online_store/test_dynamodb_online_store.py::test_dynamodb_online_store_update_tags - botocore.exceptions.ClientError: An error occurred (UnrecognizedClientException) when calling the ListTagsOfResource operation: The security token included in the request is invalid.
I'm not sure if it is related to changes that I have made. Will look into it further.