At GTC 2025, Dell unveiled an open-source connector for PowerScale designed specifically for retrieval-augmented generation (RAG) applications. This new tool facilitates greater efficiency by optimizing CPU and GPU usage on AI compute clusters, while also reducing network and storage I/O strain, leading to enhanced data processing speeds.
Significance of the PowerScale RAG Connector
In the landscape of RAG frameworks, developers often struggle to maintain their applications with the latest dataset updates. Typically, a data pipeline is established where source data is retrieved at set intervals from a storage system, such as PowerScale, which is integral to the Dell AI Factory in collaboration with NVIDIA. This pipeline is responsible for generating and updating data chunks and embeddings for the RAG application. The creation of these components is resource-intensive, consuming both CPU and GPU power.
Processing large volumes of documents, often in the range of millions and terabytes, imposes a significant load on computational, network, and storage systems, especially when duplicates are involved. The PowerScale RAG Connector alleviates this burden by intelligently determining which files have already been processed and which require updates, ultimately minimizing the amount of data needing processing. This connector is compatible with various RAG frameworks, including LangChain, as well as generic Python classes and NVIDIA’s NeMo Retriever and NIM microservices.
Functionality Overview
The PowerScale RAG connector operates in conjunction with a new feature called MetadataIQ, part of the latest software release from PowerScale. This feature periodically saves filesystem metadata to an external Elasticsearch database. The connector leverages this metadata to track previously processed files, enhancing data ingestion and processing within RAG applications.
In practice, the process follows these steps:
- Developers will use the Dell open-source Python-based RAG Connector to ingest data from PowerScale.
- The connector interacts with the Metadata Repository, which monitors new, modified, and updated files.
- After querying the database, the RAG Connector returns only the new and modified files to the RAG framework, bypassing unaltered files.
- Standard methods or the NVIDIA NeMo Retriever can be used to chunk and embed the new and modified files for processing.
- Using the NeMo Retriever’s NIM microservices, developers can quickly and accurately derive insights from extensive datasets.
This functionality allows RAG applications to exclusively process new and updated files, thereby preserving CPU, GPU, and network resources for other tasks.
Integration with NVIDIA AI Enterprise
When combined with NVIDIA AI Enterprise software, including tools like NeMo Retriever, the Dell PowerScale RAG Connector maximizes the advantages of both Dell’s technology and NVIDIA’s advanced RAG capabilities.
Accessing the PowerScale RAG Connector
Developers seeking to utilize the PowerScale RAG connector can obtain immediate access to its features and capabilities, enhancing their RAG application development.