Dell PowerScale Document Loader

Dell PowerScale is an enterprise scale out storage system that hosts industry leading OneFS filesystem that can be hosted on-prem or deployed in the cloud.

This document loader utilizes unique capabilities from PowerScale that can determine what files that have been modified since an application’s last run and only returns modified files for processing. This will eliminate the need to re-process (chunk and embed) files that have not been changed, improving the overall data ingestion workflow.

This loader requires PowerScale’s MetadataIQ feature enabled. Additional information can be found on our GitHub Repo: https://github.com/dell/powerscale-rag-connector

Overview

Integration details

Class	Package	Local	Serializable	JS support
PowerScaleDocumentLoader	powerscale-rag-connector	✅	❌	❌
PowerScaleUnstructuredLoader	powerscale-rag-connector	✅	❌	❌

Loader features

Source	Document Lazy Loading	Native Async Support
PowerScaleDocumentLoader	✅	✅
PowerScaleUnstructuredLoader	✅	✅

Setup

This document loader requires the use of a Dell PowerScale system with MetadataIQ enabled. Additional information can be found on our github page: https://github.com/dell/powerscale-rag-connector

Installation

The document loader lives in an external pip package and can be installed using standard tooling

%pip install -qU  powerscale-rag-connector

Initialization

Now we can instantiate document loader:

Generic Document Loader

Our generic document loader can be used to incrementally load all files from PowerScale in the following manner:

from powerscale_rag_connector import PowerScaleDocumentLoader

loader = PowerScaleDocumentLoader(
    es_host_url="http://elasticsearch:9200",
    es_index_name="metadataiq",
    es_api_key="your-api-key",
    folder_path="/ifs/data",
)

UnstructuredLoader Loader

Optionally, the PowerScaleUnstructuredLoader can be used to locate the changed files and automatically process the files producing elements of the source file. This is done using LangChain’s UnstructuredLoader class.

from powerscale_rag_connector import PowerScaleUnstructuredLoader

# Or load files with the Unstructured Loader
loader = PowerScaleUnstructuredLoader(
    es_host_url="http://elasticsearch:9200",
    es_index_name="metadataiq",
    es_api_key="your-api-key",
    folder_path="/ifs/data",
    # 'elements' mode splits the document into more granular chunks
    # Use 'single' mode if you want the entire document as a single chunk
    mode="elements",
)

The fields:

• es_host_url is the endpoint to MetadataIQ Elasticsearch database
• es_index_index is the name of the index where PowerScale writes it file system metadata
• es_api_key is the encoded version of your elasticsearch API key
• folder_path is the path on PowerScale to be queried for changes

Load

Internally, all code is asynchronous with PowerScale and MetadataIQ and the load and lazy load methods will return a python generator. We recommend using the lazy load function.

for doc in loader.load():
    print(doc)

[Document(page_content='' metadata={'source': '/ifs/pdfs/1994-Graph.Theoretic.Obstacles.to.Perfect.Hashing.TR0257.pdf', 'snapshot': 20834, 'change_types': ['ENTRY_ADDED']}),
Document(page_content='' metadata={'source': '/ifs/pdfs/New.sendfile-FreeBSD.20.Feb.2015.pdf', 'snapshot': 20920, 'change_types': ['ENTRY_MODIFIED']}),
Document(page_content='' metadata={'source': '/ifs/pdfs/FAST-Fast.Architecture.Sensitive.Tree.Search.on.Modern.CPUs.and.GPUs-Slides.pdf', 'snapshot': 20924, 'change_types': ['ENTRY_ADDED']})]

Returned Object

Both document loaders will keep track of what files were previously returned to your application. When called again, the document loader will only return new or modified files since your previous run.

• The metadata fields in the returned Document will return the path on PowerScale that contains the modified file. You will use this path to read the data via NFS (or S3) and process the data in your application (e.g.: create chunks and embedding).
• The source field is the path on PowerScale and not necessarily on your local system (depending on your mount strategy); OneFS expresses the entire storage system as a single tree rooted at /ifs.
• The change_types property will inform you on what change occurred since the last one - e.g.: new, modified or delete.

Your RAG application can use the information from change_types to add, update or delete entries your chunk and vector store.

When using PowerScaleUnstructuredLoader the page_content field will be filled with data from the Unstructured Loader

Lazy Load

Internally, all code is asynchronous with PowerScale and MetadataIQ and the load and lazy load methods will return a python generator. We recommend using the lazy load function.

for doc in loader.lazy_load():
    print(doc)  # do something specific with the document

The same Document is returned as the load function with all the same properties mentioned above.

Additional Examples

Additional examples and code can be found on our public github webpage: https://github.com/dell/powerscale-rag-connector/tree/main/examples that provide full working examples.

• PowerScale LangChain Document Loader - Working example of our standard document loader
• PowerScale LangChain Unstructured Loader - Working example of our standard document loader using unstructured loader for chunking and embedding
• PowerScale NVIDIA Retriever Microservice Loader - Working example of our document loader with NVIDIA NeMo Retriever microservices for chunking and embedding

API reference

For detailed documentation of all PowerScale Document Loader features and configurations head to the github page: https://github.com/dell/powerscale-rag-connector/