DS Enterprises — RAG-Based Knowledge Assistant

Mist Avinya designed and deployed a secure, serverless Retrieval-Augmented Generation (RAG) platform on AWS, enabling DS Enterprises employees to access accurate, citation-backed internal knowledge using natural language queries.

70%Reduction in information retrieval time

92%Retrieval relevance accuracy

3.8sP95 response latency

About the Customer

Business Context & Challenges

Business Context

DS Enterprises is an SMB enterprise software company with ~350 employees. Documentation was fragmented across Confluence, SharePoint, Git, and S3, slowing onboarding, engineering, and incident response.

Key Challenges

Knowledge scattered across disconnected platforms · frequent use of outdated SOPs · extended onboarding timelines · senior engineers overloaded with repetitive questions · slow retrieval impacting efficiency.

Solution

RAG Pipeline & Capabilities

Pipeline Workflow

Ingests documents from Confluence, SharePoint, Git, and S3 · semantic chunking · embeddings via Titan Embeddings V2 · vector storage in OpenSearch Serverless · semantic retrieval and re-ranking · grounded answers via Claude 3.5 Sonnet.

Platform Capabilities

Natural language conversational search · inline source citations · real-time incremental indexing · role-based access filtering · PII redaction and grounding validation · fully serverless deployment.

AWS Services

What Powers It

Amazon Bedrock

Claude 3.5 Sonnet for grounded response generation with citations.

Bedrock Knowledge Bases

Manages RAG orchestration and retrieval configuration.

OpenSearch Serverless

Vector database for scalable semantic search with metadata filtering.

Titan Embeddings V2

Semantic vector embeddings for documents and queries.

AWS Lambda

Powers ingestion, chunking, re-ranking, and API operations.

Amazon Textract

Structured text from scanned PDFs and legacy image content.

Business Outcomes

Results at a Glance

Information Retrieval Time

45–60 min/day → under 15 minutes (70% reduction)

New Hire Ramp-Up

6 weeks → 3.5 weeks (target exceeded)

Retrieval Relevance Accuracy

Manual search → 92% accuracy (target exceeded)

P95 Response Latency

N/A → 3.8 seconds (target exceeded)

Architecture

High Availability & Scalability

OpenSearch Serverless, Lambda, Bedrock, API Gateway, Cognito, and S3 operate across multiple Availability Zones with no single point of failure.

Have an internal knowledge problem to solve?

Tell us what you're working on. We'll share how we'd approach it.

Start a Conversation

← Back to GenAI & Machine Learning