Chapter 3: Security and Access Control

Security Concerns in RAG

Concern	Risk	Mitigation
Unauthorized access	Data leak	Access control
Prompt injection	Manipulation	Input validation
Data isolation	Cross-tenant	Namespaces

Document-Level Permissions

def search_with_permissions(query, user):
    """Filter results based on user permissions"""
    # Get user's allowed document IDs
    allowed_docs = get_user_permissions(user.id)
    
    # Search with filter
    results = vectorstore.similarity_search(
        query,
        filter={"doc_id": {"$in": allowed_docs}},
        k=10
    )
    
    return results

Multi-Tenant Architecture

Option 1: Namespace Isolation

# Each tenant gets a namespace
def get_tenant_retriever(tenant_id):
    return vectorstore.as_retriever(
        search_kwargs={
            "namespace": f"tenant_{tenant_id}",
            "k": 5
        }
    )

Option 2: Metadata Filtering

# Store tenant_id in metadata
def add_document(text, tenant_id):
    vectorstore.add_texts(
        [text],
        metadatas=[{"tenant_id": tenant_id}]
    )

def query_tenant(query, tenant_id):
    return vectorstore.similarity_search(
        query,
        filter={"tenant_id": tenant_id}
    )

Input Validation

def sanitize_query(query):
    """Prevent prompt injection"""
    # Remove potential injection patterns
    dangerous_patterns = [
        "ignore previous instructions",
        "disregard the above",
        "system:"
    ]
    
    for pattern in dangerous_patterns:
        if pattern.lower() in query.lower():
            raise ValueError("Invalid query")
    
    return query

Audit Logging

def log_rag_access(user_id, query, docs_accessed, response):
    """Log all RAG interactions"""
    log_entry = {
        "timestamp": datetime.utcnow(),
        "user_id": user_id,
        "query": query,
        "documents_accessed": [d.id for d in docs_accessed],
        "response_length": len(response)
    }
    audit_logger.info(json.dumps(log_entry))

Security Checklist

• Implement authentication
• Enforce document permissions
• Isolate tenant data
• Validate/sanitize inputs
• Log all access
• Encrypt sensitive data
• Regular security audits