Configuration

The provider stores credentials and model defaults in config/ai.php. The gateway reads those values from the provider on every call, so key rotation and endpoint changes are configuration operations.

Provider keys

Key Type Default Notes
driver string regolo Resolves ai.provider.regolo.
name string regolo Returned in response metadata.
key string env('REGOLO_API_KEY') Bearer token for Regolo.
url string https://api.regolo.ai/v1 Base URL without endpoint suffix.
timeout integer 60 Per-request fallback timeout.

Defaults

'defaults' => [
    'text' => env('AI_DEFAULT_TEXT', 'regolo'),
    'embeddings' => env('AI_DEFAULT_EMBEDDINGS', 'regolo'),
    'reranking' => env('AI_DEFAULT_RERANKING', 'regolo'),
],

Model aliases

Use default, cheapest, and smartest for text routing when product code should not hard-code a model id:

'models' => [
    'text' => [
        'default' => 'Llama-3.1-8B-Instruct',
        'cheapest' => 'Llama-3.1-8B-Instruct',
        'smartest' => 'Llama-3.3-70B-Instruct',
    ],
],

Embedding dimensions are part of your persistence contract. If you change models.embeddings.dimensions, rebuild downstream vector indexes that stored the previous vector length.

Configuration formula

For a retrieval pipeline with n candidate documents and top k final chunks, the reranking work can be thought of as:

cost∝n×tokens(query+document) cost \propto n \times tokens(query + document)

Keep n bounded before reranking. A common pattern is vector search for 50 candidates, rerank to 5, then send the 5 chunks to chat.