---
title: Deployment Architecture
date_updated: 2026-05-21
summary: Deployment modes (hosted | isolated-hosted | self-host), network boundaries, model egress, and connectivity controls that underpin the regulated deployments.
---

# Deployment Architecture

**Questions this doc answers**

- Which `RM_DEPLOYMENT_MODE` should I choose?
- How does Reflect Memory enforce private network boundaries and model egress policies?
- What guardrails exist for allowed model hosts, webhooks, and SSO in self-hosted deployments?

## Deployment modes table

| Mode | Ownership | Network boundary | Model egress | Public webhooks | Typical guardrails |
| --- | --- | --- | --- | --- | --- |
| hosted | Reflect Memory cloud | public | optional | allowed | Standard rate limits, telemetry, SOC 2 API |
| isolated-hosted | Dedicated runtime/DB per tenant | managed/public | configurable | restricted (per tenant) | Same infra plus tenant-level logging |
| self-host | Your VPC/air-gapped network | private | disabled by default | blocked | `RM_ALLOWED_MODEL_HOSTS`, `RM_REQUIRE_INTERNAL_MODEL_BASE_URL`, `RM_DISABLE_MODEL_EGRESS` |

## Resolve deployment config

`resolveDeploymentConfig()` centralizes the runtime flags:

- `RM_DEPLOYMENT_MODE` → `"hosted" | "isolated-hosted" | "self-host"` (default `hosted`)
- `RM_DISABLE_MODEL_EGRESS` → favors `true` in `self-host`
- `RM_REQUIRE_INTERNAL_MODEL_BASE_URL` → ensures internal models are reachable
- `RM_ALLOWED_MODEL_HOSTS` → comma-delimited whitelist enforced via `enforceModelHostPolicy`
- `RM_ALLOW_PUBLIC_WEBHOOKS` → defaults to `false` in air-gapped mode
- `RM_SSO_*` → toggles OIDC authentication; missing `JWKS`, `ISSUER`, or `AUDIENCE` ✅ fails startup

Validation is strict: self-host mode with `requireInternalModelBaseUrl` requires a non-empty `allowedModelHosts`. SSO enables per-tenant JWKS validation and email claim resolution for user lookups.

## Network boundary enforcement

Every deployment config exposes `networkBoundary` (`"public"` vs `"private"`). Self-host defaults to `"private"`. In addition:

- `requireInternalModelBaseUrl` forces every external LLM call to be proxied through your internal gateway.
- `allowedModelHosts` ensures only approved models (e.g., `llama.local`, `ollama.company`, `vicuna.private`) can be reached.
- `disableModelEgress` defaults to true for `self-host`, preventing outbound connections unless explicitly lifted.

## Pilot & upgrade flow

1. Scope call → determine `mode`, SSO/JWKS, `allowedModelHosts`, compliance requirements.  
2. Deploy pilot container/Helm chart with `RM_TENANT_ID`, `RM_SSO_*`, `RM_ALLOWED_MODEL_HOSTS`, `RM_AGENT_KEY_*`.  
3. Connect AI tools via MCP (agent keys) or REST (API key).  
4. After pilot, flip `RM_REQUIRE_INTERNAL_MODEL_BASE_URL`/`RM_DISABLE_MODEL_EGRESS` toggles as needed, then monitor audit trail + usage events for compliance.