> ## Documentation Index
> Fetch the complete documentation index at: https://docs.omnia-voice.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Omnia documentation

> Omnia measures whether your AI is good on your own traffic — calibrated judges, statistically honest comparisons — and makes model improvements safe to ship.

Omnia measures whether your AI is actually good. You capture your production
LLM traffic, grade a slice of it, and Omnia **calibrates** LLM judges against
your grades, so every number downstream (pass rates, model comparisons, canary
verdicts) is corrected for the judge's measured error and reported with a
confidence interval. On that foundation, model changes become safe to ship:
comparisons run on your own prompts, quality-gated canaries promote a new
model only when live traffic proves it, and a model trained on your traffic is
yours to run anywhere.

The measurement runs on infrastructure you already know. Omnia's gateway is an
OpenAI-compatible inference API: if your code already talks to OpenAI, point
it at Omnia's base URL and change the key; request shapes, streaming, and tool
calls stay the same. Routing through it adds one API across many open models,
caching, fallbacks, and dedicated GPU capacity — but routing is only one of
four [integration paths](/capture/overview), and two of them capture traffic
without moving your inference at all. See
[How Omnia works](/concepts/how-omnia-works).

## Base URL

One host serves everything:

```
https://gateway.omnia-voice.com/v1
```

Inference (chat, embeddings, rerank, responses, models) **and** management
(labels, criteria, evals, aliases, logs, traces, dedicated endpoints,
fine-tuning) live under the same base URL. Requests to the older management
base `https://platform.omnia-voice.com/api/v1` keep working; it's an alias,
not a migration.

Everything authenticates with the same workspace API key
(`Authorization: Bearer sk_sovereign_...`). Create one in the
[dashboard](https://platform.omnia-voice.com/dashboard) under **API keys**.

## Your first request

```bash theme={null}
curl https://gateway.omnia-voice.com/v1/chat/completions \
  -H "Authorization: Bearer $OMNIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-32B",
    "messages": [{ "role": "user", "content": "Hello!" }]
  }'
```

The response is the standard OpenAI chat-completion shape, and the `usage`
object reflects the exact tokens deducted from your wallet. See the
[Quickstart](/quickstart) for the SDK equivalents.

## Start here

<CardGroup cols={2}>
  <Card title="How Omnia works" icon="arrows-rotate" href="/concepts/how-omnia-works">
    The workflow end to end: capture, measure, improve, own.
  </Card>

  <Card title="Quickstart" icon="rocket" href="/quickstart">
    From an API key to your first response in a few minutes.
  </Card>

  <Card title="Authentication" icon="key" href="/authentication">
    How keys work, the `sk_sovereign_` format, and scoping.
  </Card>

  <Card title="API reference" icon="code" href="/api-reference/introduction">
    Every endpoint and field, interactive, generated from the API.
  </Card>

  <Card title="The gateway" icon="diagram-project" href="/concepts/gateway">
    How a request is authenticated, routed, metered, and settled.
  </Card>
</CardGroup>

## Measure and improve

<CardGroup cols={2}>
  <Card title="Grade your traffic" icon="scale-balanced" href="/judges/grading">
    Record pass/fail verdicts on real traffic: the ground truth every judge
    is measured against.
  </Card>

  <Card title="Judges & calibration" icon="gavel" href="/judges/calibration">
    Calibrate every judge against your grades: TPR, TNR, Cohen's κ, and a
    trust badge derived from confidence intervals.
  </Card>

  <Card title="Comparisons" icon="flask" href="/reference/evals">
    Test whether a cheaper model holds up on your own prompts, with pass rates
    corrected for the judge's measured error.
  </Card>

  <Card title="Self-improvement" icon="arrows-rotate" href="/reference/self-improvement">
    Use a calibrated judge as a training reward, and prove the result on
    held-out grades.
  </Card>
</CardGroup>

## Run at scale

<CardGroup cols={2}>
  <Card title="Inference" icon="plug" href="/inference/chat">
    Chat, streaming, tool calling, structured output, embeddings, and vision,
    all OpenAI-compatible.
  </Card>

  <Card title="Dedicated endpoints" icon="server" href="/dedicated/overview">
    Reserve private GPU capacity for a model, billed per GPU-hour while running
    instead of per token.
  </Card>

  <Card title="Fine-tuning" icon="wand-magic-sparkles" href="/fine-tuning/overview">
    Train a model on your own data, then serve it on a dedicated endpoint and
    call it like any other model.
  </Card>

  <Card title="Billing" icon="wallet" href="/concepts/billing">
    Prepaid wallet, per-token and per-GPU-hour metering, holds, and auto-reload.
  </Card>
</CardGroup>

## How a request flows

Your app talks to Omnia exactly like it talks to OpenAI. Behind the endpoint,
Omnia authenticates the key to a workspace, routes the request to the model
across multiple clouds, meters the exact usage, and settles it against your
wallet in real time.

```mermaid theme={null}
flowchart LR
    A[Your app] -->|OpenAI-compatible request| B(Omnia gateway)
    B -->|key -> workspace| C[Auth + wallet gate]
    C --> D[Model routing]
    D --> E[Inference]
    E -->|response + usage| B
    B -->|metered per token| F[(Wallet)]
```