Resources

Frequently asked questions

Short answers to the questions IT, platform and security teams ask most often. Each module's page has the details.

Setup and integration

  • Do we need to rewrite our applications? No: just a new endpoint URL: Your application connects to the OpenAI-compatible chat endpoint with a project API key. If you turn on per-user limits, add the X-FFM-User header to every request.
  • Where is BaseLLM installed? Your infrastructure: Docker or Helm: Install it on a single server with Docker, or on Kubernetes and OpenShift with the Helm chart. During a rolling update, the old version keeps running until the new one is ready.
  • How is the license verified? Signed, tied to your installation: The license is signed in the portal. The application checks the signature every time it starts, and the signature is valid only for that installation. Copied, modified or expired licenses are rejected.
  • What happens to my license if we change servers? You get a new license for the new server: With Docker, the enterprise license is tied to the server it is installed on; on Kubernetes and OpenShift, it is tied to the whole cluster, not to a single server. Scaling replicas or adding and removing servers in the cluster doesn't affect it. When you move to another server, or to a new or standby cluster, give the new installation's fingerprint to the portal and get a new license.

Data and security

  • Does our data leave the organization? Only to your model provider: If you use an external provider, the text sent to the model goes to that provider; BaseLLM scans it before it leaves. User identities aren't passed to the provider, and personal data isn't stored in raw form.
  • How is sensitive data protected? Rule-based scanning and masking: BaseLLM recognizes types such as Turkish national ID numbers (TCKN), phone numbers, email addresses, IBANs, card numbers, tax IDs, passport numbers, license plates, IP addresses and dates of birth. Secret scanning is on in every project and, by default, blocks with 403 secret_blocked.
  • What does the Audit Log record? Who did what, and the outcome: Each entry records the actor, correlation_id, action, path, outcome, IP hash and user-agent; prompts, model responses and raw API keys are never written. Each entry carries the fingerprint of the previous one, so any change or deletion shows up in chain verification.
  • Can we export to a SIEM? 6 channels, 9 SIEM products: The audit, usage, pii, injection, policy and operational_error channels are sent as JSON or CEF every 60 seconds by default (15–3,600 s). Prompts and model responses are not sent.

Operations and cost

  • What happens if a provider fails? Can use a fallback model: When fallback is enabled for the project, timeouts, connection failures, 429, 5xx and rejected keys move the request to the next model; invalid requests don't. For streaming responses, fallback happens only before the first token is sent.
  • What happens when a project's budget runs out? Block or Warn only: In Block mode, requests are rejected with 429 budget_exceeded. In Warn only mode, they go through and are logged. The alert threshold defaults to 80%, and the alert is logged in the Audit Log and sent to your SIEM.
  • How much latency does the gateway add? p99 1.73–1.76 ms: Measured in a test environment, the added time is p50 0.29–0.30 ms and p99 1.73–1.76 ms. This covers only the time BaseLLM adds; it doesn't include the model's response time.