Connect Large Language Models
Configure OpenAI- or Anthropic-compatible API endpoints for ALAP asynchronous threat reasoning and self-learning rules.
CheeseWAF’s ALAP (AI Large-Language-Model Auto Pilot) engine integrates with industry-standard LLM APIs to perform out-of-band semantic reasoning on suspicious and borderline traffic.
In the Web management console, navigate to the AI section and configure the following parameters:
| Setting | Configuration Key | Description & Examples |
|---|---|---|
| Enable AI | ai.enabled | Master toggle for ALAP functionality (boolean) |
| Provider | ai.provider | API protocol dialect: openai or anthropic |
| Base URL | ai.api_base | API Base URL (e.g., https://api.openai.com/v1) |
| Allow Private Endpoint | ai.allow_private_api_base | Allows connecting to private/local LLM gateways (Ollama, LocalAI; defaults to false) |
| API Key | ai.api_key | Secret access token / API key provided by your LLM vendor |
| Model Name | ai.model | Model identifier (e.g., gpt-4o-mini, claude-3-5-sonnet-latest) |
| Auto-Agreement | sites[].waf.semantic_policy.auto_agree | When enabled, threats evaluated as high or critical are automatically persisted as defensive rules |
Legacy configuration migration: The old ai.base_url and ai.auto_agree keys are not read by the current release. Move ai.base_url to ai.api_base, and move the approval setting to each site’s sites[].waf.semantic_policy.auto_agree; save the configuration and reload/restart the service as required.
Best Practices & Security Guidelines
- Enforce Asynchronous Execution: Always maintain
ai.async: truein your configuration to guarantee that the Data Plane never blocks real-time traffic while waiting for model inference. - Connection Testing: Before enabling rules, click Test Connection in the console to validate network routing, DNS resolution, and credential authentication.
- Gradual Rollout of Auto-Agreement: Keep
sites[].waf.semantic_policy.auto_agreedisabled during initial deployment. Review samples manually in the threat queue for 1–2 weeks to verify evaluation fidelity before enabling automated rule generation.
For details on assistant tool approvals, review queues, and automated self-learning, see ALAP & Review Queue.