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

# Provider Configuration

> Configure OpenAI-compatible providers independently for each aCMF role.

# Provider Configuration

aCMF supports separate OpenAI-compatible providers for:

* Adjudicator
* Context Enhancer
* Cortex
* Embeddings

## Provider kinds

* `openai_compatible`
* `stub`

`stub` is useful for tests and local development. `openai_compatible` expects a standard OpenAI-style API with `base_url`, API key, and model name.

For a complete field-by-field environment reference, including which variables are required and when, see [Environment Reference](/guides/environment-reference).

## Recommended models

| Role             | Recommended models                                                                                      |
| ---------------- | ------------------------------------------------------------------------------------------------------- |
| Adjudicator      | `openai/gpt-5.4` or `google/gemini-3.1-flash-lite-preview`                                              |
| Context Enhancer | `openai/gpt-5.4`, `google/gemini-3.1-flash-lite-preview`, or `inception/mercury-2` for fast token speed |
| Cortex           | `openai/gpt-5.4` or `google/gemini-3.1-flash-lite-preview`                                              |
| Embedding        | `openai/text-embedding-3-large` or `openai/text-embedding-3-small`                                      |

Recommended starting point:

* use `openai/gpt-5.4` for adjudicator, context enhancer, and cortex if you want one consistent default
* use `inception/mercury-2` only for the Context Enhancer if you want faster token speed on read-time synthesis
* use `openai/text-embedding-3-large` when retrieval quality matters most, or `openai/text-embedding-3-small` when cost and vector size matter more

## Resolution rules

* Each role first reads its own `PROVIDER`, `BASE_URL`, `API_KEY`, `MODEL`, and `TIMEOUT_SECONDS`.
* If a role does not define `BASE_URL`, `API_KEY`, or `TIMEOUT_SECONDS`, aCMF falls back to the global OpenAI-compatible defaults.
* If a role is set to `stub`, network credentials are not required for that role.

## Global defaults

These settings are shared when a role does not define its own override:

```env theme={null}
ACMF_OPENAI_DEFAULT_BASE_URL=https://api.openai.com/v1
ACMF_OPENAI_DEFAULT_API_KEY=
ACMF_OPENAI_DEFAULT_TIMEOUT_SECONDS=30
```

## Adjudicator

```env theme={null}
ACMF_ADJUDICATOR_PROVIDER=openai_compatible
ACMF_ADJUDICATOR_BASE_URL=
ACMF_ADJUDICATOR_API_KEY=
ACMF_ADJUDICATOR_MODEL=openai/gpt-5.4
ACMF_ADJUDICATOR_TIMEOUT_SECONDS=
```

## Context Enhancer

```env theme={null}
ACMF_CONTEXT_ENHANCER_PROVIDER=openai_compatible
ACMF_CONTEXT_ENHANCER_BASE_URL=
ACMF_CONTEXT_ENHANCER_API_KEY=
ACMF_CONTEXT_ENHANCER_MODEL=openai/gpt-5.4
ACMF_CONTEXT_ENHANCER_TIMEOUT_SECONDS=
```

## Cortex

```env theme={null}
ACMF_CORTEX_PROVIDER=openai_compatible
ACMF_CORTEX_BASE_URL=
ACMF_CORTEX_API_KEY=
ACMF_CORTEX_MODEL=openai/gpt-5.4
ACMF_CORTEX_TIMEOUT_SECONDS=
```

## Embeddings

```env theme={null}
ACMF_EMBEDDING_PROVIDER=openai_compatible
ACMF_EMBEDDING_BASE_URL=
ACMF_EMBEDDING_API_KEY=
ACMF_EMBEDDING_MODEL=openai/text-embedding-3-large
ACMF_EMBEDDING_DIMENSIONS=3072
ACMF_EMBEDDING_TIMEOUT_SECONDS=
```

## Example mixed setup

```env theme={null}
ACMF_OPENAI_DEFAULT_BASE_URL=https://api.openai.com/v1
ACMF_OPENAI_DEFAULT_API_KEY=sk-default

ACMF_ADJUDICATOR_PROVIDER=openai_compatible
ACMF_ADJUDICATOR_MODEL=openai/gpt-5.4

ACMF_CONTEXT_ENHANCER_PROVIDER=openai_compatible
ACMF_CONTEXT_ENHANCER_BASE_URL=https://my-router.example/v1
ACMF_CONTEXT_ENHANCER_API_KEY=router-key
ACMF_CONTEXT_ENHANCER_MODEL=inception/mercury-2

ACMF_CORTEX_PROVIDER=openai_compatible
ACMF_CORTEX_MODEL=google/gemini-3.1-flash-lite-preview

ACMF_EMBEDDING_PROVIDER=openai_compatible
ACMF_EMBEDDING_MODEL=openai/text-embedding-3-large
ACMF_EMBEDDING_DIMENSIONS=3072
```

## Recommendation

Keep providers independent, but start by using the same endpoint for all roles unless you have a clear routing reason not to. The important boundary is that each role can choose its own model and credentials when needed.