RAG SYSTEMS

RAG Platform for Retail Knowledge

Centralised multi-tenant RAG core turning documentation chaos into conversational clarity

Challenge

•Knowledge scattered across 60+ unstructured sources (wikis, docs, tickets, drives)
•Employees relying on 'ask a friend' instead of search; low trust in existing tools
•Each new assistant was a one-off build, tightly coupled to a specific LLM provider
•No clean model for multi-tenant isolation (per BU / country / role)

•Designed an end-to-end RAG platform with multi-tenant vector store (namespaces per org / BU / country)
•Implemented hybrid dense + lexical retrieval to handle noisy enterprise text
•Built pluggable LLM abstraction to swap OpenAI, Gemini, or local models without rewrites
•Created ingestion pipelines for 60+ sources with cleaning, splitting, enrichment and change-capture
•Exposed a unified retrieval + generation API consumed by multiple assistants (FAQ bots, search copilots, support tools)
•Applied hexagonal / ports-and-adapters architecture so new data sources or models mean new adapters, not new cores
•Instrumented everything with observability: retrieval hit-rate, latency, errors, and assistant usage

•One shared RAG core powering several assistants, instead of a separate stack per use case
•Higher trust in answers and fewer 'where is X?' tickets, based on internal feedback
•Faster onboarding of new teams: plug in data sources, configure tenant, deploy an assistant
•Reduced lock-in risk by abstracting LLM providers behind a stable interface

Dedicated evaluation layer de-risking LLM deployments with measurable precision and recall metrics