Powering Cross-Department Knowledge Search for Paragon with a Retrieval-Augmented Generation System on AWS

Industry: Professional Services Region: AWS Europe (Ireland) AWS Services Used: Amazon S3, Amazon OpenSearch Serverless, Amazon Bedrock (Anthropic Claude 3.5 Sonnet, Amazon Titan Embeddings), AWS Lambda, Amazon DynamoDB, Amazon API Gateway, Amazon CloudFront, AWS CloudFormation, Amazon CloudWatch

Executive Summary

Paragon partnered with VArrow Technologies to replace fragmented, third-party AI tools with a secure, fully managed Retrieval-Augmented Generation (RAG) system built natively on AWS. The solution gives sales, HR, and accounting teams fast, contextual answers drawn directly from Paragon’s own document library — more than 10,000 files synced weekly from Google Drive — without ever sending sensitive company data to an external AI provider for training. VArrow designed and delivered the system end-to-end: from the storage foundation that anchors the entire pipeline, through the retrieval and generation layer powered by Amazon Bedrock, to a serverless architecture built for cost efficiency and scale.

About the Customer

Paragon relies on a growing library of internal documents — spreadsheets, PDFs, and policy files — spread across sales, HR, and accounting functions. As that library grew past 10,000 documents, finding the right information quickly became harder, and teams had started turning to general-purpose AI chat tools to search and summarize company content — tools that offered no guarantee about how that data was being used or retained.

The Challenge

Paragon needed a way to make its internal knowledge base searchable and useful in natural language, without compromising on data security or control. Specifically, the solution had to:

• Replace ad hoc use of external AI tools with a secure system that keeps all company data inside Paragon’s own AWS environment
• Deliver fast, accurate, context-aware answers grounded in Paragon’s actual documents — not generic AI knowledge
• Serve three distinct departments with different content and different needs, without requiring separate systems
• Scale cost-effectively as usage and document volume grow, without paying for idle infrastructure

The Solution: A RAG System Built on a Solid Storage Foundation

VArrow designed a Retrieval-Augmented Generation (RAG) architecture — a pattern that grounds an AI model’s answers in an organization’s own content, rather than relying on the model’s general training data alone. Every part of that pattern depends on getting the underlying data foundation right, and that’s where the design started.

Amazon S3: the foundation the whole system is built on

At the center of the architecture is Amazon S3, which serves as the single, durable source of truth for Paragon’s document library. Every document synced weekly from Google Drive lands in S3 first — nothing is processed, indexed, or made searchable until it’s safely and durably stored there. S3 also hosts the system’s web UI as static assets, keeping the delivery layer simple and cost-efficient.

This “storage-first” design matters because everything downstream depends on it: – Ingestion reads directly from S3 to extract and chunk document content before embedding. – Object tagging was implemented across the document bucket to categorize files (by department and source), which in turn drives S3 lifecycle policies for cost-efficient long-term storage. – Access pattern analysis informed the storage class and lifecycle configuration, ensuring frequently retrieved documents stay performant while older, less-accessed content transitions to lower-cost storage tiers automatically. – Because S3 is durable and independent of the search index sitting on top of it, Paragon’s document library remains intact and recoverable even as the AI/search layer evolves — the retrieval and generation components are, in effect, a smart lens on top of a storage layer that isn’t going anywhere.

Retrieval: Amazon OpenSearch Serverless

Once documents are in S3, content is extracted, chunked, and converted into vector embeddings, then indexed in Amazon OpenSearch Serverless. This gives Paragon fast semantic search — finding documents based on meaning and context rather than exact keyword matches — without the operational overhead of managing a search cluster.

Generation: Amazon Bedrock with Anthropic’s Claude 3.5 Sonnet

For the generative layer, VArrow selected Amazon Bedrock, using Amazon Titan for embeddings generation and Anthropic’s Claude 3.5 Sonnet as the response-generation model. Claude 3.5 Sonnet was chosen for its strong balance of reasoning quality and response speed — well suited to producing accurate, context-grounded answers to department-specific business questions without the latency cost of larger models. Because both the embedding and generation steps run inside Amazon Bedrock, Paragon’s document content never leaves its own AWS account or gets used to train an external model — directly addressing the data security requirement that motivated the project in the first place.

Tying it together: a serverless, cost-efficient pipeline

The rest of the system was built serverless end-to-end: – AWS Lambda handles both the weekly document ingestion pipeline and real-time query processing. – Amazon DynamoDB stores conversation history with automatic capacity management. – Amazon API Gateway and Amazon CloudFront deliver a secure REST API and a responsive web interface. – AWS CloudFormation provisions the entire environment as Infrastructure as Code, making the deployment repeatable and auditable. – Amazon CloudWatch provides full observability — tracking query latency, ingestion success rates, and system health continuously.

This approach means Paragon pays only for what it uses, with no idle infrastructure sitting between queries.

Results

• Unified knowledge access across sales, HR, and accounting — one system, department-aware responses
• 10,000+ documents made instantly searchable via natural language, with weekly automated ingestion from Google Drive
• Zero external data exposure — all document processing, embedding, and generation happens inside Paragon’s own AWS account, with no data used for third-party AI training
• Cost-efficient, serverless design — infrastructure scales with actual usage rather than fixed capacity
• Continuous monitoring of query performance, ingestion health, and system availability via CloudWatch, giving Paragon’s team ongoing visibility rather than a one-time delivery

Why It Matters

For Paragon, the win wasn’t just faster search — it was replacing an uncontrolled, external AI habit with a purpose-built system that keeps sensitive company information exactly where it belongs. By anchoring the architecture on a durable Amazon S3 storage foundation and layering retrieval and generation on top with Amazon OpenSearch Serverless and Amazon Bedrock’s Claude 3.5 Sonnet, VArrow delivered a system that’s both genuinely useful day-to-day and built to scale without re-architecture as Paragon’s document library keeps growing.

VArrow Technologies is an AWS Partner specializing in cloud migration, AI/ML solutions, and managed services. To learn how VArrow can help design a secure, AWS-native RAG solution for your organization, contact us.