Structured Semantic RAG with Access Control

Retrieval-Augmented Generation (RAG) improves language models by letting them pull in external context — but traditional implementations fall short for enterprise use. Aegis extends this pattern with Structured Semantic RAG: a secure, metadata-aware, graph-enabled retrieval system that supports fine-grained access control.

This guide explains what makes Structured Semantic RAG (SSRAG) different, why it’s needed, and how it enables enterprise-grade retrieval for intelligent agents.

Why Simple Semantic Search Falls Short

Imagine building an assistant for a multi-tenant healthcare platform. A physician asks:

“Summarize the patient’s respiratory history.”

Standard RAG will:

• Embed the query
• Run approximate nearest neighbor search
• Return the top-k closest document chunks

Problems:

• May return data from other patients or departments
• Doesn’t filter by user role or visibility
• Doesn’t understand document structure or relationships

In enterprise settings, this leads to:

• Data leakage across tenants
• Inconsistent or irrelevant answers
• Lack of control and traceability

What SSRAG Adds

Structured Semantic RAG enhances traditional RAG with:

1. Structured Metadata Filtering

Every document and chunk in Aegis is tagged with metadata like:

• tenant_id, document_type, access_visibility, tags, owner_id

At query time, we filter based on:

• User’s tenant and role
• Required tags or document types
• Visibility constraints

Example:

Only return chunks from physician-notes tagged respiratory, authored within the user’s department, and marked private.

2. Semantic Graph Relationships (GraphRAG)

Chunks are not isolated — they’re linked semantically:

• “Patient History” → “Treatment Plan” → “Follow-up Notes”
• “Clause” → “Cited Case” → “Commentary”

Aegis supports traversing this graph:

• Multi-hop reasoning
• Re-ranking based on neighborhood
• Expansion from anchor concepts to supporting context

Example:

A legal agent asks for “termination clause context.” SSRAG expands from the clause to related cases and commentary, returning a richer, more precise summary.

3. Access-Aware Retrieval

Access control is not applied post-retrieval — it’s embedded in the retrieval process:

• Chunk access policies are enforced in the Vespa query layer
• Users only see what they’re permitted to see, at index time

Example:

An intern from Hospital A cannot retrieve notes from Hospital B, even if embeddings are close.

Why Vespa?

Vespa is uniquely suited to power SSRAG:

• Hybrid Retrieval: Lexical (BM25) + Semantic (ANN)
• Structured Filtering: First-class support for metadata-based filters
• Ranking Expressions: Re-rank using recency, semantic similarity, access score, etc.
• High Performance: Scales across large, multi-tenant deployments

Other vector databases lack this combination of structured filters, hybrid scoring, and runtime ranking flexibility.

Use Cases Enabled

Multi-Tenant AI Workflows

Guarantee data boundaries across organizations with chunk-level access enforcement.

Scoped Retrieval with Reasoning

Traverse related context across a semantic document graph.

Domain-Specific Control

Retrieve based on structured metadata: department, doc type, access, version.

What’s Next

• Learn about document and chunk schemas
• Explore the Vespa setup
• Read about access control models

SSRAG turns your RAG pipeline into a permission-aware, structured knowledge base — unlocking the next generation of enterprise-grade AI agents.