Pebblo enables developers to safely load data and promote their Gen AI app to deployment without worrying about the organization’s compliance and security requirements. The project identifies semantic topics and entities found in the loaded data and summarizes them on the UI or a PDF report.

Pebblo has two components.

1. Pebblo Safe DocumentLoader for Langchain
2. Pebblo Server

This document describes how to augment your existing Langchain DocumentLoader with Pebblo Safe DocumentLoader to get deep data visibility on the types of Topics and Entities ingested into the Gen-AI Langchain application. For details onPebblo Server see thispebblo server document.

Pebblo Safeloader enables safe data ingestion for LangchainDocumentLoader. This is done by wrapping the document loader call withPebblo Safe DocumentLoader.

Note: To configure pebblo server on some url other that pebblo's default (localhost:8000) url, put the correct URL inPEBBLO_CLASSIFIER_URL env variable. This is configurable using theclassifier_url keyword argument as well. Ref:server-configurations

How to Pebblo enable Document Loading?​

Assume a Langchain RAG application snippet usingCSVLoader to read a CSV document for inference.

Here is the snippet of Document loading usingCSVLoader.

from langchain_community.document_loadersimport CSVLoader

loader= CSVLoader("data/corp_sens_data.csv")
documents= loader.load()
print(documents)

The Pebblo SafeLoader can be enabled with few lines of code change to the above snippet.

from langchain_community.document_loadersimport CSVLoader, PebbloSafeLoader

loader= PebbloSafeLoader(
    CSVLoader("data/corp_sens_data.csv"),
    name="acme-corp-rag-1",# App name (Mandatory)
    owner="Joe Smith",# Owner (Optional)
    description="Support productivity RAG application",# Description (Optional)
)
documents= loader.load()
print(documents)

Send semantic topics and identities to Pebblo cloud server​

To send semantic data to pebblo-cloud, pass api-key to PebbloSafeLoader as an argument or alternatively, put the api-key inPEBBLO_API_KEY environment variable.

from langchain_community.document_loadersimport CSVLoader, PebbloSafeLoader

loader= PebbloSafeLoader(
    CSVLoader("data/corp_sens_data.csv"),
    name="acme-corp-rag-1",# App name (Mandatory)
    owner="Joe Smith",# Owner (Optional)
    description="Support productivity RAG application",# Description (Optional)
    api_key="my-api-key",# API key (Optional, can be set in the environment variable PEBBLO_API_KEY)
)
documents= loader.load()
print(documents)

Add semantic topics and identities to loaded metadata​

To add semantic topics and sematic entities to metadata of loaded documents, set load_semantic to True as an argument or alternatively, define a new environment variablePEBBLO_LOAD_SEMANTIC, and setting it to True.

from langchain_community.document_loadersimport CSVLoader, PebbloSafeLoader

loader= PebbloSafeLoader(
    CSVLoader("data/corp_sens_data.csv"),
    name="acme-corp-rag-1",# App name (Mandatory)
    owner="Joe Smith",# Owner (Optional)
    description="Support productivity RAG application",# Description (Optional)
    api_key="my-api-key",# API key (Optional, can be set in the environment variable PEBBLO_API_KEY)
    load_semantic=True,# Load semantic data (Optional, default is False, can be set in the environment variable PEBBLO_LOAD_SEMANTIC)
)
documents= loader.load()
print(documents[0].metadata)

Anonymize the snippets to redact all PII details​

Setanonymize_snippets toTrue to anonymize all personally identifiable information (PII) from the snippets going into VectorDB and the generated reports.

Note: ThePebblo Entity Classifier effectively identifies personally identifiable information (PII) and is continuously evolving. While its recall is not yet 100%, it is steadily improving.For more details, please refer to thePebblo Entity Classifier docs

from langchain_community.document_loadersimport CSVLoader, PebbloSafeLoader

loader= PebbloSafeLoader(
    CSVLoader("data/corp_sens_data.csv"),
    name="acme-corp-rag-1",# App name (Mandatory)
    owner="Joe Smith",# Owner (Optional)
    description="Support productivity RAG application",# Description (Optional)
    anonymize_snippets=True,# Whether to anonymize entities in the PDF Report (Optional, default=False)
)
documents= loader.load()
print(documents[0].metadata)

Related​

• Document loaderconceptual guide
• Document loaderhow-to guides