Skip to content

Run and maintain LQ.AI

  • operator
  • evaluator

This is the operator’s namespace: getting LQ.AI running, keeping it running, and knowing what to do when it doesn’t. Docker Compose is the reference installation, tested end to end on this site; start there, then come back for the week-two work once it’s up.

Five install paths have a recipe today. Pick the one that matches how you plan to run LQ.AI — each page names what it covers and what it doesn’t. Windows and corporate-cloud-specific deployment (running the stack itself on Azure, AWS, or GCP infrastructure) are not covered by a recipe yet; the gateway’s Azure OpenAI provider adapter is a model provider you can point the stack at, not a guide to hosting the app on Azure.

  • macOS desktop app — a signed, notarized launcher for Apple Silicon. No terminal, no cloned repository.
  • Docker Compose — the reference deployment: clone, edit .env, docker compose up -d.
  • Kubernetes via Helm — a chart for the core chat services; it omits some background workers the Compose setup runs. Read the page before you rely on it for more than that.
  • Reverse proxy and TLS — putting a certificate and a stable URL in front of either install above.
  • Air-gapped / local-only inference — the same Compose stack with local inference instead of a cloud key, and what’s actually been verified about its network behavior.

Docker Compose, cloud provider keys (Mode 1)

Status recipe published Recipe Read the recipe

Notes
Checked against README.md.

Docker Compose, local-only inference (Mode 2, –profile local)

Status recipe published Recipe Read the recipe

Notes
the compose stack brings this up; no CI job in this repository yet measures its network behavior — see the page for the egress-inventory gap. Checked against README.md.

Kubernetes via the in-repo Helm chart

Status recipe published Recipe Read the recipe

Notes
the chart has no Deployment for the ingest-worker or arq-worker background workers — document ingestion and queued jobs do not run on it yet (PRD DE-327). See the page before you rely on it for anything beyond chat. Checked against deploy/helm/lq-ai/.

Reverse proxy + TLS, public domain (Caddy, Traefik, or nginx with a public certificate)

Status recipe published Recipe Read the recipe

Notes
Caddy, Traefik and nginx overlays under deploy/reverse-proxy/, labelled by their README as reference configurations rather than a supported production distribution; not deployed against a public domain for this site. Checked against deploy/reverse-proxy/README.md.

Windows host

Status not tried Recipe none

Notes
not documented in the repository as of the checked commit.

Hosted on a corporate cloud VM or Kubernetes cluster (Azure, AWS, GCP)

Status not tried Recipe none

Notes
not documented in the repository as of the checked commit — distinct from the Azure OpenAI inference provider adapter, which is shipped.

If your shape isn’t on this list, treat it as untried: nothing in the repository documents it as of the checked commit. The table is rendered at build time from _data/supported-shapes.yaml, so what you read here and what the repository records are the same list.

Assign responsibility for updates, backups, storage, access, and leaked keys before you rely on a deployment. Once LQ.AI is up: hardware sizing, backup and restore, upgrading, rotating a leaked provider key, and triage when something is wrong. Moving or uninstalling and the troubleshooting index round these out. The recipes index covers a private Tailscale address, a separate Ollama computer, and monitoring — each in the same format: problem, context, steps, verify, and the refusal errors you might hit.

Headless boot is the one page on running the stack without the web UI — acknowledged, not supported, no compatibility promised.