APEMIND / SECURITY & ARCHITECTURE
ApeMind security, data residency & access architecture
Security reviews for enterprise RAG platforms depend on concrete boundaries, not broad compliance adjectives. Every row below is verified directly against ApeMind deployment, permissions, governance, and connection documentation—including explicit current limitations.
Verifiable security & architecture facts
| Area | Verified behavior | Explicit boundary / operator responsibility | Source |
|---|---|---|---|
| Deployment Topologies | Offered as SaaS (apemind.ai) or two official private deployment architectures: single-VM Docker Compose and Kubernetes + KubeBlocks Enterprise (3+ servers with multi-replica Frontend, API, and Index Workers). | Docker Compose operates within a single-host fault domain. Kubernetes HA requires KubeBlocks Enterprise for managed database lifecycle, backups, and failover. | Deployment architecture |
| Data Storage Locations | Raw files are stored in object storage (standalone RustFS chart by default on new private installs, or legacy KubeBlocks MinIO when explicitly enabled); metadata in PostgreSQL; knowledge graph in Graph PostgreSQL; vectors in Qdrant; full-text index in Elasticsearch; queues and cache in Redis. | Operators own infrastructure disks, backup retention, restoration drills, and cluster access policies in private deployments. | Topology & stores |
| Network & Outbound Boundaries | In private deployment, core application and data services stay inside your VPC/cluster. Model endpoints, parsers (MinerU, Mistral OCR, LlamaParse), web search/read, and external Connections are explicitly configured per deployment or organization. | Private deployment does not automatically make all integrations offline. Audit every configured LLM, embedding, rerank, OCR, web-access, and MCP/HTTP endpoint before handling sensitive documents. | Models & routing |
| Dual-Plane Role Isolation | Site roles (admin, rw, ro) govern platform operations, users, quotas, and rate limits. Organization roles (Owner, Admin, Editor, Viewer, plus custom action-level roles) govern knowledge bases, agents, datasets, and connections. | Site admins do not automatically gain access to organization business data; reading an organization collection still requires membership and matching action permission. | Permissions model |
| Default Deny & Existence Masking | All entry points (Web UI, REST API, Agents, MCP, Share links) share one authorization kernel evaluated on every real read. Unlisted actions are denied by default. | Org members lacking an action receive explicit denial; callers unrelated to a resource receive "not found" so unauthorized users cannot probe resource existence. | Access resolution |
| API Keys & Machine Identities | Every API key is pinned to a single workspace at creation and never floats when the holder switches workspaces. Supports Normal keys (30/90/365-day expiry), Managed agent keys (lifecycle-bound, protected from manual deletion), and Service Account keys (non-human org credentials, plaintext shown once). | Data API keys cannot invoke platform admin endpoints. Org governance lists all keys pinned to the org and allows revoking unprotected keys. | Machine identities |
| External Connection Credentials | Connection credentials and upstream OAuth tokens are write-only (never returned in API responses, logs, audit entries, or errors) and stored in a dedicated table separate from non-secret config. Changing target host, URL, or auth mode clears stored credentials automatically. | Documented limitation: connection credentials and upstream OAuth tokens are stored in plaintext in PostgreSQL (not encrypted at rest). Database backups include them and must be protected and encrypted at the storage/backup layer. | Connection security |
| Database Connection Safeguards | PostgreSQL and MySQL connections apply read-only SQL AST filtering (allowing SELECT / read-only CTE / UNION; blocking DML, DDL, multi-statement, COPY, EXPLAIN, SET, system catalogs, pg_sleep, dblink, LOAD_FILE) and execute inside read-only transactions capped at 200 rows, 256 KiB, and 30s. | Application guards do not replace database-side least privilege: always create a dedicated read-only database user. Grant connection.create only to trusted roles because test-connection probes internal DB ports. | DB connection limits |
| Protocol Gates (MCP & A2A) | MCP Server (/mcp/) enforces layered workspace_access, collection_access, and web_access gates and never falls back to browser cookies or server env keys. A2A 1.0 Agent Cards (/v1/a2a/bots/{bot_id}/agent-card.json) are private and require a workspace-bound API key. | ApeMind does not expose a public /.well-known/agent-card.json endpoint; unauthorized A2A callers receive HTTP 404. | A2A & MCP gates |
| Quotas, Rate Limits & Audit | Supports organization quotas (collections, total docs, docs per collection, bots), system and per-user rate-limit policies with optimistic revision locking (409 REVISION_CONFLICT), 1–168h rejection records, and API/tool audit logs. | Quotas are enforced per organization, not per individual user. Rejection logs explain HTTP 429 events and complement—rather than replace—the API/MCP audit ledger. | Governance & audit |
| Open-Source Auditability | The open-source core codebase (ApeRAG) is publicly available on GitHub for code inspection; ApeMind Enterprise builds commercial governance, enterprise connectors, and managed private delivery on top. | Verify whether a specific enterprise feature lives in open-source ApeRAG or the commercial ApeMind Enterprise distribution during procurement. | Product overview |
Pre-production security verification checklist
- Map every outbound dependency (LLM, embedding, rerank, OCR/parser, web search, external MCP/HTTP) and verify network egress rules.
- Enable storage-level and backup encryption for PostgreSQL, object storage, and volumes—especially because Connection credentials are stored in plaintext in PostgreSQL.
- Test cross-organization isolation with two separate user identities and workspace-bound API keys; verify unrelated callers receive "not found".
- Confirm site-admin accounts cannot read organization knowledge bases unless explicitly added as org members.
- Verify backup and restore procedures for PostgreSQL, Graph PG, Qdrant, Elasticsearch, and RustFS before onboarding production data.
Security & architecture FAQ
Can a platform site admin read every organization’s private knowledge base?
No. ApeMind separates Site roles (admin, rw, ro) from Organization roles (Owner, Admin, Editor, Viewer, and custom roles). Site admins manage users, quotas, and deployment configuration, but do not automatically gain read access to organization knowledge bases, documents, or agents without org membership and action permissions.
Does private deployment guarantee zero outbound network traffic?
Not automatically. Private deployment keeps ApeMind core services and data stores in your infrastructure, but outbound traffic still occurs if you configure external cloud LLM APIs, cloud OCR parsers, web search, or remote MCP/HTTP connections. Point those dependencies to internal endpoints or disable them to run fully offline.
How are external Connection credentials stored and protected?
Credentials and upstream OAuth tokens are write-only in the API and UI, excluded from logs and audit records, stored in a separate database table, and automatically wiped if target hosts or auth modes change. However, they are currently stored in plaintext in PostgreSQL rather than encrypted at rest, so database storage and backups must be tightly access-controlled and encrypted.
How does ApeMind prevent agents from reading unauthorized documents?
Every retrieval and tool invocation passes through the same authorization kernel evaluated at read time. Prompt hints or tool arguments cannot widen authorization, and API keys remain pinned to the single workspace selected when they were created.
Is ApeMind open source and auditable?
The core RAG engine, ApeRAG, is publicly open source on GitHub (github.com/apecloud/ApeRAG) for inspection. ApeMind Enterprise is the commercial enterprise distribution supporting SaaS and private Docker Compose or Kubernetes + KubeBlocks Enterprise deployments.
Related architecture & evaluation pages
Review deployment sizing and architecture
Estimate compute, storage, and database resources for Docker Compose or Kubernetes + KubeBlocks Enterprise before scheduling a pilot.
Read sizing reference