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

# Connect your own LLM provider or gateway

> Have Clarifeye call your own model provider account or LLM gateway, route calls through your gateway, and choose which models your organization uses.

By default, Clarifeye calls the model providers with its own accounts. On a dedicated instance, you can have Clarifeye call your own provider account or your own LLM gateway instead. Your quota, billing and data-processing terms then apply.

The Clarifeye team sets up the connection when we configure your instance. Contact us, then send the details listed below for your provider through a secure channel.

> [!WARNING]
>
> Do not send API keys, service-account keys or client secrets by email.

## Supported providers and gateways

| Provider or gateway | What we need from you |
| - | - |
| OpenAI | API key |
| Anthropic | API key |
| Mistral | API key |
| Google Gemini Enterprise Agent Platform (formerly Vertex AI) | Service-account key (JSON file), GCP project ID, region |
| Azure OpenAI | Endpoint URL, API key, deployment names |
| OpenRouter | API key |
| LiteLLM gateway | Gateway base URL, and either an API key or OAuth client credentials |

## OpenAI, Anthropic, Mistral

Create an API key in the provider's console and send it to us. You do not need to send an endpoint: Clarifeye uses the provider's public API.

## Google Gemini Enterprise Agent Platform (formerly Vertex AI)

Use this option to keep every call in your own GCP project.

We need:

* A **service-account key** in JSON format. The service account must have the `roles/aiplatform.user` role, and the `aiplatform.googleapis.com` API must be enabled in the project.
* The **GCP project ID**.
* The **region**, for example `eu` to keep processing in the EU.

## Azure OpenAI

We need:

* The **endpoint** of your Azure resource, for example `https://<resource>.openai.azure.com`. Send the base endpoint, not the full Target URI of a deployment.
* An **API key** for that resource.
* The **deployment names** of the models you want us to use.

## OpenRouter

Create an API key in your OpenRouter account. You do not need to send an endpoint.

## LiteLLM gateway

Use this option if you run your own LiteLLM proxy in front of your providers.

We need:

* The **base URL** of the gateway, for example `https://llm-gateway.example.com/v1`. It must be reachable from the Clarifeye instance.
* One way to authenticate:
  * a **static API key** (a LiteLLM virtual key), or
  * **OAuth client credentials**, if your gateway sits behind Microsoft Entra ID or another identity provider: token URL, client ID, client secret and scope. Clarifeye requests and renews the token itself.
* The **model names** exposed by your gateway (as listed by `GET <base URL>/models`).

Once the gateway is connected, you can send all calls through it, or only some models. See [Route calls through your LLM gateway](#route-calls-through-your-llm-gateway).

## Route calls through your LLM gateway

> [!NOTE]
>
> Gateway routing is not available on all plans. Contact us to activate it for your organization.

Gateway routing sends LLM and agentic calls through your [LiteLLM gateway](#litellm-gateway) instead of calling providers (OpenAI, Anthropic, Gemini, Mistral) directly. This lets you centralize provider keys, spend, logging and access policy in one place, while Clarifeye keeps working exactly as before.

There are two ways to use it:

* **Every model**: Clarifeye routes all LLM calls through the gateway, and infers the provider from the model name. Ask us to enable this.
* **Per model**: you opt in one model at a time with the `lite-llm-gateway/` prefix. Models you leave unprefixed keep their default route, so you can move one tier, one extractor or one Playground agent onto the gateway without touching the rest.

### What can be routed

Any model that Clarifeye lets you name can go through the gateway, including:

* The [model overrides](#model-overrides) of your organization.
* The **Playground** assistant (chat, agentic tool use, and playbook execution).
* **Extractors**: object extraction, tag extraction, and document-tag extraction.

### Set the model name

To route a model through the gateway, prefix its name with `lite-llm-gateway/` in the model field:

```text theme={null}
lite-llm-gateway/<provider>/<model>
```

Clarifeye strips the `lite-llm-gateway/` prefix and passes the remainder (`<provider>/<model>`) to the gateway verbatim, so it must match a model your gateway is configured to serve. A prefixed model always goes through the gateway, whatever the default route.

For reasoning models, add a reasoning-effort suffix after a colon (`:low`, `:medium`, `:high`), exactly as you would without the gateway:

```text theme={null}
lite-llm-gateway/<provider>/<model>:<effort>
```

### Examples

| Provider | Direct model name | Through the gateway |
| - | - | - |
| OpenAI | `gpt-5.4` | `lite-llm-gateway/openai/gpt-5.4` |
| Anthropic | `anthropic/claude-opus-4-8` | `lite-llm-gateway/anthropic/claude-opus-4-8` |
| Gemini | `gemini/gemini-2.5-pro` | `lite-llm-gateway/gemini/gemini-2.5-pro` |
| Mistral | `mistral/mistral-large-latest` | `lite-llm-gateway/mistral/mistral-large-latest` |

With a reasoning-effort suffix:

```text theme={null}
lite-llm-gateway/openai/gpt-5:medium
lite-llm-gateway/openai/gpt-5:high
```

> [!TIP]
>
> Whitespace around the value is ignored, so a stray leading or trailing space won't accidentally bypass gateway routing.

## Model overrides

> [!NOTE]
>
> Only organization admins can set model overrides.

Model overrides replace the platform's default model for a tier with a model of your choice, for every store in the organization:

1. Open your organization and click **Settings** in the sidebar.
2. Under **Model overrides**, enter a model name next to each tier you want to change. The placeholder shows the platform default for that tier.
3. Click **Save**.

The tiers are `STRONG_MODEL`, `STRONG_HIGH_REASONING`, `STRONG_LOW_REASONING`, `MEDIUM_MODEL`, `LIGHT_MODEL`, and `EMBEDDING_MODEL`. Leave a tier blank to keep the platform default.

To send a tier through your gateway, enter its gateway model name, as described in [Set the model name](#set-the-model-name).

## Good to know

* **Choosing models.** Clarifeye uses models at several levels (strong, medium, light, embedding). We map each level to a model from your provider. Organization admins can then change the model for each level with [model overrides](#model-overrides).
* **Embeddings.** The embedding model must be chosen before the first document is added. Changing it afterwards means processing every document again.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.