Inference API

OpenAI-compatible inference under /v1. Auth is an org API key (ad_…) in the Authorization: Bearer header. The standard OpenAI SDKs work with base_url set to https://lorivo.dev/v1.

Model ids

• qwen3-4b: the base model.
• qwen3-4b:<adapter>: an adapter from your org. Only adapters with status LOADED resolve; anything else returns 404 Unknown model.

Adapters are namespaced by org: two orgs can each have an adapter called my-lora and each key sees its own.

curl https://lorivo.dev/v1/models \
  -H "Authorization: Bearer ad_…"

{
  "object": "list",
  "data": [
    { "id": "qwen3-4b", "object": "model", "owned_by": "adapters" },
    { "id": "qwen3-4b:my-lora", "object": "model", "owned_by": "acme-ai" }
  ]
}

curl https://lorivo.dev/v1/chat/completions \
  -H "Authorization: Bearer ad_…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-4b:my-lora",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'

{
  "id": "chatcmpl-…",
  "object": "chat.completion",
  "model": "qwen3-4b:my-lora",
  "choices": [
    {
      "index": 0,
      "message": { "role": "assistant", "content": "Hello! How can I help?" },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 12, "completion_tokens": 9, "total_tokens": 21 }
}

curl https://lorivo.dev/v1/completions \
  -H "Authorization: Bearer ad_…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen3-4b:my-lora",
    "prompt": "Summarize the plot of Dune.",
    "max_tokens": 128
  }'

curl https://lorivo.dev/v1/responses \
  -H "Authorization: Bearer ad_…" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen3-4b:my-lora","input":"Hello!","stream":true}'

Streaming

All three completion endpoints stream server-sent events when stream: true. Use curl -N to see events as they arrive.

Usage and limits

• Every request is logged against your org (prompt + completion tokens); it shows up on the /usage dashboard page.
• Request bodies are capped at 16 MiB.
• If a node restart drops a warm adapter, the first request self-heals by reloading it and retrying once.

Common errors