* Remove outdated documentation files: Quick Start Guide, Apache AGE Analysis, and Scratchpad.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.
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
- Replace
get_tenant_context_optionalwithget_tenant_contextfor all data endpoints - Remove admin bypass in
tenant_service.py - Make
user_idrequired in RAG manager
Priority 2: Defense-in-Depth
- Enable PostgreSQL RLS for all tenant-scoped tables
- Validate workspace format to prevent injection
- Add audit logging for cross-tenant access attempts
Priority 3: Operational Improvements
- Add cache invalidation on tenant/KB deletion
- Implement rate limiting per tenant
- 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
- Create Tenant A and KB
- Upload document to Tenant A
- Attempt to list documents with Tenant B headers
- Expected: Empty list for Tenant B
Scenario 2: Missing Tenant Header
- Send document list request without X-Tenant-ID header
- Expected: 400 Bad Request OR empty response (not global data)
Scenario 3: Query Isolation
- Insert data into Tenant A KB
- Query with Tenant B context
- 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:
- Optional tenant context fallback to global RAG
- Admin user bypass in access control
- No database-level security (RLS)
Recommended Next Steps
- 🔴 Immediate: Fix SEC-001 and SEC-002 before production
- 🟡 Short-term: Enable PostgreSQL RLS
- 🟢 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) |
C. Related Documentation
docs/0001-multi-tenant-architecture.mddocs/0002-multi-tenant-visual-reference.mddocs/0003-multi-tenant-documentation-index.mdMULTITENANT_IMPLEMENTATION_SUMMARY.md