LightRAG/docs/archives/action_plan/multitenant-audit/06-final-audit-report.md
Raphael MANSUY 2b292d4924
docs: Enterprise Edition & Multi-tenancy attribution (#5)
* Remove outdated documentation files: Quick Start Guide, Apache AGE Analysis, and Scratchpad.

* Add multi-tenant testing strategy and ADR index documentation

- Introduced ADR 008 detailing the multi-tenant testing strategy for the ./starter environment, covering compatibility and multi-tenant modes, testing scenarios, and implementation details.
- Created a comprehensive ADR index (README.md) summarizing all architecture decision records related to the multi-tenant implementation, including purpose, key sections, and reading paths for different roles.

* feat(docs): Add comprehensive multi-tenancy guide and README for LightRAG Enterprise

- Introduced `0008-multi-tenancy.md` detailing multi-tenancy architecture, key concepts, roles, permissions, configuration, and API endpoints.
- Created `README.md` as the main documentation index, outlining features, quick start, system overview, and deployment options.
- Documented the LightRAG architecture, storage backends, LLM integrations, and query modes.
- Established a task log (`2025-01-21-lightrag-documentation-log.md`) summarizing documentation creation actions, decisions, and insights.
2025-12-04 18:09:15 +08:00

12 KiB

Multi-Tenant Implementation Audit - Final Report

Date: November 29, 2025
Auditor: Automated Audit Process
Version: 1.0


Executive Summary

This audit reviewed the multi-tenant implementation of the LightRAG stack from Web UI through REST API to Storage layer. The implementation uses a workspace-based isolation model where workspace = tenant_id/kb_id.

Overall Assessment: ⚠️ FUNCTIONAL WITH CONCERNS

The multi-tenant implementation is functionally correct but has several security and architectural concerns that should be addressed before production deployment.


Architecture Overview

┌─────────────────────────────────────────────────────────────────┐
│                         Web UI (React)                          │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────────────┐  │
│  │ TenantStore  │  │ KBSelector   │  │ Axios Interceptor    │  │
│  │ (Zustand)    │  │              │  │ X-Tenant-ID Header   │  │
│  └──────────────┘  └──────────────┘  └──────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
                              │
                    HTTP Headers (X-Tenant-ID, X-KB-ID)
                              │
┌─────────────────────────────────────────────────────────────────┐
│                      REST API (FastAPI)                         │
│  ┌──────────────────┐  ┌────────────────┐  ┌─────────────────┐ │
│  │ get_tenant_      │  │ TenantContext  │  │ TenantRAG       │ │
│  │ context()        │──│ (dataclass)    │──│ Manager         │ │
│  │ dependency       │  │                │  │ (per-tenant RAG)│ │
│  └──────────────────┘  └────────────────┘  └─────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
                              │
                    workspace = tenant_id/kb_id
                              │
┌─────────────────────────────────────────────────────────────────┐
│                     Storage Layer                               │
│  ┌──────────────────────────────────────────────────────────┐  │
│  │ PostgreSQL Tables with workspace column                   │  │
│  │ PRIMARY KEY (workspace, id)                               │  │
│  │                                                           │  │
│  │ Tables: LIGHTRAG_DOC_FULL, LIGHTRAG_DOC_CHUNKS,          │  │
│  │         LIGHTRAG_VDB_*, LIGHTRAG_TENANTS, etc.           │  │
│  └──────────────────────────────────────────────────────────┘  │
│  ┌─────────────────┐  ┌─────────────────┐  ┌────────────────┐  │
│  │ Redis           │  │ Apache AGE      │  │ Vector Store   │  │
│  │ (key prefix)    │  │ (graph per ws)  │  │ (ws filter)    │  │
│  └─────────────────┘  └─────────────────┘  └────────────────┘  │
└─────────────────────────────────────────────────────────────────┘

Test Results Summary

Unit Tests: ALL PASSED

Test Suite Tests Result
test_multitenant_e2e.py 32 PASSED
test_tenant_security.py 11 PASSED
test_multi_tenant_backends.py 37 PASSED
test_tenant_storage_phase3.py 22 PASSED
TOTAL 102 ALL PASSED

Key Verifications

  • Cross-tenant data access prevention
  • Knowledge base isolation within tenants
  • Role-based permission enforcement
  • Composite key uniqueness
  • Redis namespace isolation
  • Vector ID uniqueness per tenant

Security Findings

Critical Findings

ID Severity Component Finding
SEC-001 🔴 HIGH API get_tenant_context_optional() allows global RAG fallback
SEC-002 🔴 HIGH API Admin user bypass in tenant_service.py
SEC-003 🟡 MEDIUM API user_id parameter optional in RAG manager
SEC-004 🟡 MEDIUM Storage No PostgreSQL Row-Level Security (RLS)
SEC-005 🟢 LOW Storage Tenant context set for logging only
SEC-006 🟢 LOW WebUI localStorage used for token storage

SEC-001: Global RAG Fallback (HIGH)

Location: lightrag/api/routers/document_routes.py, query_routes.py

Issue: Using get_tenant_context_optional allows requests without tenant headers to use global RAG:

async def get_tenant_rag(
    tenant_context: Optional[TenantContext] = Depends(get_tenant_context_optional)
) -> LightRAG:
    if rag_manager and tenant_context:
        return await rag_manager.get_rag_instance(...)
    return rag  # Falls back to global RAG!

Risk: Data leakage if global RAG contains multi-tenant data.

Recommendation: Use get_tenant_context (required) for all data endpoints in multi-tenant mode.

SEC-002: Admin User Bypass (HIGH)

Location: lightrag/services/tenant_service.py

Issue:

if user_id.lower() == "admin":
    return True  # Bypass all access checks

Risk: Any user with username "admin" can access any tenant's data.

Recommendation: Implement proper super-admin role with separate authentication.

SEC-003: Optional User ID (MEDIUM)

Location: lightrag/tenant_rag_manager.py

Issue: user_id parameter is optional, allowing access without authentication check.

Recommendation: Deprecate unauthenticated access path.

SEC-004: No Row-Level Security (MEDIUM)

Location: PostgreSQL Tables

Issue: Tenant isolation relies entirely on application-level workspace filtering. No database-level enforcement.

Recommendation: Enable PostgreSQL RLS as defense-in-depth:

ALTER TABLE LIGHTRAG_DOC_FULL ENABLE ROW LEVEL SECURITY;
CREATE POLICY tenant_isolation ON LIGHTRAG_DOC_FULL
    USING (workspace = current_setting('app.current_workspace'));

Architectural Findings

ARCH-001: Workspace Encoding

Pattern: workspace = "{tenant_id}/{kb_id}"

Strengths:

  • Simple, hierarchical namespace
  • Clear data ownership
  • Compatible with file-based storage

Concerns:

  • No explicit validation of workspace format
  • Potential collision if tenant IDs contain /

ARCH-002: Context Propagation

Pattern: Python ContextVar for tenant context

tenant_id_var: ContextVar[Optional[str]] = ContextVar("tenant_id")

Current Use: Set for PostgreSQL session logging, but NOT used for security.

Recommendation: Use context variable for RLS enforcement.

ARCH-003: RAG Instance Caching

Pattern: LRU cache keyed by (tenant_id, kb_id)

Strengths:

  • Memory efficient
  • Fast subsequent access

Concerns:

  • Cache size limit may cause thrashing with many tenants
  • No cache invalidation on tenant deletion

Web UI Findings

UI-001: Tenant State Management

Location: lightrag_webui/src/stores/tenant.ts

Status: Well implemented

  • Zustand store with persistence
  • Automatic header injection via Axios interceptor
  • Console logging for debugging

UI-002: Document State Clearing

Location: lightrag_webui/src/features/DocumentManager.tsx

Recent Fix: State clearing on tenant/KB change

useEffect(() => {
  setCurrentPageDocs([]);
  setDocs(null);
  setStatusCounts({ all: 0 });
  // ...
}, [selectedTenant?.tenant_id, selectedKB?.kb_id]);

Status: Fixed (per task log dated 2025-02-25)

UI-003: Token Storage

Finding: Authentication token stored in localStorage

Risk: XSS vulnerability could expose tokens

Recommendation: Use httpOnly cookies for production


Recommendations

Priority 1: Critical Security Fixes

  1. Replace get_tenant_context_optional with get_tenant_context for all data endpoints
  2. Remove admin bypass in tenant_service.py
  3. Make user_id required in RAG manager

Priority 2: Defense-in-Depth

  1. Enable PostgreSQL RLS for all tenant-scoped tables
  2. Validate workspace format to prevent injection
  3. Add audit logging for cross-tenant access attempts

Priority 3: Operational Improvements

  1. Add cache invalidation on tenant/KB deletion
  2. Implement rate limiting per tenant
  3. Add metrics/monitoring for tenant isolation verification

Test Protocol for Manual Verification

Prerequisites

# Start test databases
docker compose -f docker-compose.test-db.yml up -d

# Verify containers
docker ps --filter "name=lightrag-audit"

Test Scenarios

Scenario 1: Cross-Tenant Data Access Prevention

  1. Create Tenant A and KB
  2. Upload document to Tenant A
  3. Attempt to list documents with Tenant B headers
  4. Expected: Empty list for Tenant B

Scenario 2: Missing Tenant Header

  1. Send document list request without X-Tenant-ID header
  2. Expected: 400 Bad Request OR empty response (not global data)

Scenario 3: Query Isolation

  1. Insert data into Tenant A KB
  2. Query with Tenant B context
  3. Expected: Query returns no results from Tenant A

Test Script

# Run API isolation tests
python docs/action_plan/multitenant-audit/scripts/test_api_isolation.py

Conclusion

The LightRAG multi-tenant implementation provides functional tenant isolation through:

  • Workspace-based table partitioning
  • Per-tenant RAG instance management
  • Header-based tenant context extraction
  • Role-based permission model

However, the implementation has security gaps that must be addressed:

  1. Optional tenant context fallback to global RAG
  2. Admin user bypass in access control
  3. No database-level security (RLS)
  1. 🔴 Immediate: Fix SEC-001 and SEC-002 before production
  2. 🟡 Short-term: Enable PostgreSQL RLS
  3. 🟢 Long-term: Add comprehensive audit logging and monitoring

Appendices

A. Files Reviewed

Layer Files
Web UI tenant.ts, client.ts, DocumentManager.tsx
API dependencies.py, document_routes.py, query_routes.py, tenant_routes.py
Service tenant_service.py, tenant_rag_manager.py
Storage postgres_impl.py, utils_context.py, namespace.py
Tests test_multitenant_e2e.py, test_tenant_security.py, test_multi_tenant_backends.py

B. Test Environment

Component Version/Config
Python 3.12.12
PostgreSQL pgvector/pgvector:pg16
Redis 7-alpine
OS macOS (arm64)
  • docs/0001-multi-tenant-architecture.md
  • docs/0002-multi-tenant-visual-reference.md
  • docs/0003-multi-tenant-documentation-index.md
  • MULTITENANT_IMPLEMENTATION_SUMMARY.md