Support Azure OpenAI for self-hosted onboarding

[alex-on-ai]

Summary

• add a shared AI provider that uses Azure OpenAI when Azure env vars are configured, while preserving AI Gateway/OpenAI behavior for default installs
• make onboarding completion resilient when Trigger background jobs cannot be enqueued immediately
• throttle and retry embedding upserts so low-quota Azure OpenAI deployments can finish linkage work
• extend the risk/vendor linkage Trigger task max duration for slower self-hosted runs

Why

Self-hosted deployments that do not have AI Gateway configured currently fail during onboarding/vendor/risk generation. On Azure OpenAI, linkage can also fail on embedding rate limits because task/risk/vendor embedding batches run concurrently and the task hits the default 5 minute duration.

Validation

• Verified equivalent patch on a self-hosted Azure VM with Azure OpenAI, Trigger dev runner, Postgres, Redis/Vector, and MinIO.
• On the VM, onboarding generated 18 policies, 6 vendors, 5 risks, 35 task embeddings, 5 risk embeddings, 6 vendor embeddings, 40 risk-task links, and 48 vendor-task links.
• Ran targeted ESLint on changed app files: no errors, existing warnings only.
• Ran app typecheck; it currently fails on unrelated existing repo issues such as missing @trycompai/company / @trycompai/billing modules and pre-existing test fixture type errors.
• Ran secret scan over this diff: no env secrets/tokens detected.

Notes

This is opened as a draft because it was derived from a self-hosted POC patch. It intentionally excludes deployment-only VM files such as Caddy config, Docker overrides, .env backups, and generated Prisma artifacts.

---

Summary by cubic

Adds Azure OpenAI support for self-hosted installs and hardens onboarding/linkage under low quotas and missing workers. Default installs keep using AI Gateway/OpenAI without changes.

• New Features

  • Introduced a shared AI provider that prefers Azure when AZURE_* env vars are set, falling back to AI Gateway or OpenAI.
  • Updated chat and embedding calls to use the provider; supports Azure deployment-based models.
  • Added @ai-sdk/azure.

• Bug Fixes

  • Onboarding completes even if Trigger jobs can’t be enqueued; background tasks are best-effort.
  • Embedding upserts now batch and retry with backoff; configurable via env. Also serialized task/risk/vendor upserts to reduce spikes.
  • Increased linkage task max duration to 20 minutes for slower self-hosted runs.

^{Written for commit 43f07b4. Summary will update on new commits.}

[vercel]

@alex-on-ai is attempting to deploy a commit to the Comp AI Team on Vercel.

A member of the Team first needs to authorize it.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}