Architecture

The eLightInfo RAG Architecture

A production-grade RAG system is far more than a vector database and an LLM. Our architecture addresses ingestion, retrieval quality, security isolation, and evaluation at enterprise scale.

Layer 1

Document Ingestion Pipeline

Automated connectors for PDF, Word, Excel, PowerPoint, HTML, SharePoint, Confluence, Salesforce Knowledge, and custom APIs. Intelligent chunking strategies (semantic, recursive, agentic) with metadata preservation for rich filtering during retrieval.

Layer 2

Embedding & Vector Store

High-quality embedding models (text-embedding-3-large, Cohere embed-v3) with optimized chunk sizes per document type. Stored in your preferred vector database: Pinecone, Weaviate, pgvector (PostgreSQL), or Amazon OpenSearch. All data remains in your cloud account.

Layer 3

Hybrid Retrieval Strategy

Combine dense vector search with BM25 keyword search (hybrid retrieval) for better recall. Cross-encoder reranking to ensure the most relevant chunks top the list. Metadata filtering (by date, department, document type, classification level) for precision.

Layer 4

Secure Generation Layer

LLM generates responses grounded in retrieved context with mandatory citation. Hallucination detection via faithfulness scoring. User-level access controls ensure responses only reference documents the querying user is authorized to see.

Layer 5

Evaluation & Continuous Improvement

RAGAS-based automated evaluation: answer faithfulness, context relevancy, and answer relevancy tracked continuously. Human evaluation loop for quality sampling. Retrieval quality dashboards for ongoing tuning.

Supported Document Sources

Document Files

PDF, Word (.docx), Excel (.xlsx), PowerPoint, plain text, Markdown

Microsoft 365

SharePoint Online, Teams channels, OneDrive, OneNote

Atlassian Suite

Confluence spaces and pages, Jira tickets and comments

CRM & Custom APIs

Salesforce Knowledge, ServiceNow, custom REST/GraphQL APIs

Supported Vector Stores

Pinecone

Managed, serverless vector database with sub-millisecond query latency

Weaviate

Open-source with built-in hybrid search and multi-tenant isolation

pgvector (PostgreSQL)

Vectors in your existing Postgres database — no new infrastructure

Amazon OpenSearch

AWS-managed hybrid search with k-NN and BM25 in a single service

Security

Multi-Tenant RAG with Document-Level Access Control

In enterprise environments, different users and teams should only be able to query documents they are authorized to access. Our RAG implementation enforces document-level access controls at query time — not just at ingestion time.

Users authenticated via your IdP (Okta, Azure AD, Cognito) receive a permission context that filters the vector search to their authorized document set before any retrieval occurs. This prevents cross-tenant data leakage even when multiple business units share the same RAG infrastructure.

HIPAA-ready SOC 2 Data residency controls Audit logging PII redaction

All data stays in your cloud account

No document content ever leaves your environment.

End-to-end encryption

Encryption at rest and in transit with customer-managed keys.

Unlock Your Organization's Knowledge

A focused 2-week RAG proof-of-concept connects your most critical knowledge source to an intelligent query interface. See the value before committing to a full rollout.

Start an Enterprise RAG PoC