Local AI Provider

CaseSolo supports firm-controlled local AI routing for Ollama, vLLM, and OpenAI-compatible local servers. The launch contract is policy-first: every local/cloud routing decision must be deterministic, privilege-aware, residency-aware, and auditable without leaking secrets.

Settings Shape

Local AI is stored in Firm.settings.localAI:

{
  "enabled": true,
  "type": "ollama",
  "baseUrl": "http://gpu-node.internal:11434",
  "model": "llama3.1:70b",
  "apiKey": "optional-local-secret",
  "maxConcurrency": 4,
  "timeoutMs": 120000
}

Data residency is stored in Firm.settings.dataResidency:

{
  "region": "us",
  "mode": "local_preferred",
  "forceLocal": false,
  "allowCloudFallback": true
}

Supported mode values are cloud_allowed, local_preferred, and local_only. region: "on-prem" or forceLocal: true always normalizes to local_only.

Policy Rules

• Only ADMIN and SUPER_ADMIN can manage local AI settings or manually request a local provider.
• Any authorized user can be routed to local AI when firm policy requires it for privileged, HIPAA-sensitive, or local-only residency work.
• privilegedOnlyLocal: true routes privileged matters to the configured local provider and blocks cloud fallback if no local provider is available.
• local_preferred uses local AI when available and falls back to cloud only when allowCloudFallback is not false.
• Local connection testing records an ai.provider_selection_tested activity with a redacted providerSelection audit payload.
• Settings changes record an ai.local_settings_updated activity with redacted provider metadata and a provider-selection preview.

Implementation Points

• parseLocalAIConfig() normalizes Ollama/vLLM defaults and ignores disabled providers.
• parseDataResidencyConfig() normalizes residency controls into one mode.
• resolveAIProviderSelection() returns an allowed/blocked decision with a secret-safe ProviderSelectionAuditPayload.
• getComplianceAwareProvider() wraps the decision and returns either LocalAIProvider or the cloud default provider.
• redactLocalAIConfig() exposes provider type, model, endpoint host, and hasApiKey, never raw API keys.

Use the audit payload directly in Activity.metadata.providerSelection for AI execution, provider tests, and settings changes.