Skip to content

Deployment Guides

Koji doesn't ship its own model. It extracts against whatever endpoint you configure in the Model Providers screen of the dashboard (Settings → Model Providers). Five paths are supported today. Pick the one that matches how your organisation buys and runs inference.

Guide For One-line positioning
OpenAI Teams that just want it to work The fastest path. A key, a model, done.
Azure OpenAI Regulated buyers Same models as OpenAI, inside your Azure tenancy and DPA.
AWS Bedrock AWS-native shops Claude, Titan, Llama through your existing AWS account.
Anthropic (direct) Claude-first users Direct access to the latest Claude snapshots without a cloud middleman.
On-prem / self-hosted Air-gapped or cost-sensitive Point Koji at vLLM, TGI, or Ollama over HTTP.

All five use the same stored-endpoint flow: credentials are entered once in the UI, encrypted at rest, and referenced by pipeline steps via an endpoint ID. Per-pipeline model overrides still work — see Configuration.

Which one should I pick?

  • Just starting out? Use OpenAI with gpt-4o-mini. Cheapest, fastest to set up, good-enough accuracy for most schemas.
  • Compliance / data residency matters? Use Azure OpenAI — the data processing agreement covers extraction calls under the same terms as the rest of your Azure workload.
  • Already committed to AWS? Use Bedrock. You avoid onboarding a new vendor and stay inside one bill.
  • Want the latest Claude first? Use Anthropic direct. New model snapshots land there before they cycle through Bedrock or Azure.
  • Air-gapped network or volume-driven cost problem? Use on-prem. Any OpenAI-compatible inference server (vLLM is the most production-ready) will work.

Mixing providers is fine and common — you can pin the map step to a cheap OpenAI model and run the final extract step against Bedrock Claude, for example.

Parse / OCR is BYO too

Models handle extraction; parse / OCR handles turning scanned PDFs and images into text. That cost is yours to own as well. Configure an OCR vendor key in Project settings → Parse Endpoints and the per-page parse cost stays on your bill — Mistral OCR, Azure Document Intelligence, Google Document AI, or AWS Textract. See Parse / OCR Providers. Digital PDFs parse for free in-process; with no parse endpoint configured the built-in default engine is used.