Configuration

After publishing the configuration file with php artisan vendor:publish --tag=instructor-config, you will find it at config/instructor.php. This file controls every aspect of the package, from LLM provider connections and extraction behavior to HTTP transport, logging, event bridging, code-agent execution, native-agent runtime boundaries, and response caching. This is Laravel-native configuration. The Laravel integration reads config('instructor.*') through Laravel’s config repository and converts those arrays into typed runtime config objects internally. It does not ask the standalone packages/config YAML loader to parse config/instructor.php.

Default Connection

return [
    'default' => env('INSTRUCTOR_CONNECTION', 'openai'),
];
// @doctest id="79f3"

This determines which LLM connection is used when you call a facade without specifying one explicitly. You can override it at runtime with ->connection('name') on any facade, or by passing an LLMConfig object via ->fromConfig(...).

Connections

Configure multiple LLM provider connections. Each connection defines its driver, API credentials, default model, and token limits. You can define as many connections as you need and switch between them at runtime.

return [
    'connections' => [
        'openai' => [
            'driver' => 'openai',
            'api_url' => env('OPENAI_API_URL', 'https://api.openai.com/v1'),
            'api_key' => env('OPENAI_API_KEY'),
            'organization' => env('OPENAI_ORGANIZATION'),
            'model' => env('OPENAI_MODEL', 'gpt-4o-mini'),
            'max_tokens' => env('OPENAI_MAX_TOKENS', 4096),
        ],

        'anthropic' => [
            'driver' => 'anthropic',
            'api_url' => env('ANTHROPIC_API_URL', 'https://api.anthropic.com/v1'),
            'api_key' => env('ANTHROPIC_API_KEY'),
            'model' => env('ANTHROPIC_MODEL', 'claude-sonnet-4-20250514'),
            'max_tokens' => env('ANTHROPIC_MAX_TOKENS', 4096),
        ],

        'azure' => [
            'driver' => 'azure',
            'api_key' => env('AZURE_OPENAI_API_KEY'),
            'resource_name' => env('AZURE_OPENAI_RESOURCE'),
            'deployment_id' => env('AZURE_OPENAI_DEPLOYMENT'),
            'api_version' => env('AZURE_OPENAI_API_VERSION', '2024-08-01-preview'),
            'model' => env('AZURE_OPENAI_MODEL', 'gpt-4o-mini'),
            'max_tokens' => env('AZURE_OPENAI_MAX_TOKENS', 4096),
        ],

        'gemini' => [
            'driver' => 'gemini',
            'api_url' => env('GEMINI_API_URL', 'https://generativelanguage.googleapis.com/v1beta'),
            'api_key' => env('GEMINI_API_KEY'),
            'model' => env('GEMINI_MODEL', 'gemini-2.0-flash'),
            'max_tokens' => env('GEMINI_MAX_TOKENS', 4096),
        ],

        'ollama' => [
            'driver' => 'ollama',
            'api_url' => env('OLLAMA_API_URL', 'http://localhost:11434/v1'),
            'api_key' => env('OLLAMA_API_KEY', 'ollama'),
            'model' => env('OLLAMA_MODEL', 'llama3.2'),
            'max_tokens' => env('OLLAMA_MAX_TOKENS', 4096),
        ],
    ],
];
// @doctest id="f7d4"

Supported Drivers

Driver	Provider	Description
`openai`	OpenAI	GPT-4, GPT-4o, GPT-4o-mini
`anthropic`	Anthropic	Claude 3, Claude 3.5, Claude 4
`azure`	Azure OpenAI	Azure-hosted OpenAI models
`gemini`	Google	Gemini 1.5, Gemini 2.0
`mistral`	Mistral AI	Mistral, Mixtral models
`groq`	Groq	Fast inference with Llama, Mixtral
`cohere`	Cohere	Command models
`deepseek`	DeepSeek	DeepSeek models
`ollama`	Ollama	Local open-source models
`perplexity`	Perplexity	Perplexity models

Adding a Custom Connection

Any OpenAI-compatible API can be used by setting the openai driver and pointing api_url to your endpoint. Extra keys beyond the standard set (driver, api_url, api_key, endpoint, model, max_tokens, options) are automatically merged into the options array and forwarded with each request.

return [
    'connections' => [
        // ... existing connections
        'my-custom' => [
            'driver' => 'openai', // Use OpenAI-compatible API
            'api_url' => 'https://my-custom-api.com/v1',
            'api_key' => env('MY_CUSTOM_API_KEY'),
            'model' => 'custom-model',
            'max_tokens' => 4096,
        ],
    ],
];
// @doctest id="78dd"

Embeddings Connections

Configure embedding model connections separately from inference connections. The embeddings section has its own default key and connection definitions.

return [
    'embeddings' => [
        'default' => env('INSTRUCTOR_EMBEDDINGS_CONNECTION', 'openai'),
        'connections' => [
            'openai' => [
                'driver' => 'openai',
                'api_url' => env('OPENAI_API_URL', 'https://api.openai.com/v1'),
                'api_key' => env('OPENAI_API_KEY'),
                'model' => env('OPENAI_EMBEDDINGS_MODEL', 'text-embedding-3-small'),
                'dimensions' => env('OPENAI_EMBEDDINGS_DIMENSIONS', 1536),
            ],
            'ollama' => [
                'driver' => 'ollama',
                'api_url' => env('OLLAMA_API_URL', 'http://localhost:11434/v1'),
                'api_key' => env('OLLAMA_API_KEY', 'ollama'),
                'model' => env('OLLAMA_EMBEDDINGS_MODEL', 'nomic-embed-text'),
                'dimensions' => env('OLLAMA_EMBEDDINGS_DIMENSIONS', 768),
            ],
        ],
    ],
];
// @doctest id="3c93"

Extraction Settings

Configure defaults for structured output extraction. These values apply to every StructuredOutput call unless overridden at runtime.

return [
    'extraction' => [
        // Output mode: json_schema, json, tools, md_json
        'output_mode' => env('INSTRUCTOR_OUTPUT_MODE', 'json_schema'),
        // Maximum retry attempts when validation fails
        'max_retries' => env('INSTRUCTOR_MAX_RETRIES', 2),
        // Prompt template for retry attempts
        'retry_prompt' => 'The response did not pass validation. Please fix the following errors and try again: {errors}',
    ],
];
// @doctest id="a83e"

Output Modes

The output mode controls how the package instructs the LLM to produce structured output. Different providers have varying levels of support for each mode.

Mode	Description	Best For
`json_schema`	Uses JSON Schema for structured output	Most reliable; recommended for OpenAI
`json`	Simple JSON mode without schema enforcement	Fallback for models that lack schema support
`tools`	Uses tool/function calling to extract structured data	Alternative approach; good cross-provider support
`md_json`	Markdown-wrapped JSON	Useful for Gemini and similar models

HTTP Client Settings

Configure the underlying HTTP transport. The Laravel package ships with its own LaravelDriver that wraps Laravel’s HTTP client (Illuminate\Http\Client\Factory), which means Http::fake() works transparently in your tests.

return [
    'http' => [
        // Driver: 'laravel' uses Laravel's HTTP client (enables Http::fake())
        'driver' => env('INSTRUCTOR_HTTP_DRIVER', 'laravel'),
        // Request timeout in seconds
        'timeout' => env('INSTRUCTOR_HTTP_TIMEOUT', 120),
        // Connection timeout in seconds
        'connect_timeout' => env('INSTRUCTOR_HTTP_CONNECT_TIMEOUT', 30),
    ],
];
// @doctest id="71ff"

The service provider binds Cognesy\Http\Contracts\CanSendHttpRequests to the Laravel-backed HTTP transport. All higher-level services (Inference, Embeddings, StructuredOutput) depend on that contract, ensuring consistent HTTP behavior across the entire package.

Logging Settings

The package includes a logging pipeline that enriches log entries with Laravel request context (request ID, authenticated user, route, URL) automatically.

return [
    'logging' => [
        // Enable/disable logging
        'enabled' => env('INSTRUCTOR_LOGGING_ENABLED', true),
        // Log channel (must exist in config/logging.php)
        'channel' => env('INSTRUCTOR_LOG_CHANNEL', 'stack'),
        // Minimum log level
        'level' => env('INSTRUCTOR_LOG_LEVEL', 'warning'),
        // Logging preset: default, production, or custom
        'preset' => env('INSTRUCTOR_LOGGING_PRESET', 'production'),
        // Events to exclude from logging
        'exclude_events' => [
            Cognesy\Http\Events\DebugRequestBodyUsed::class,
            Cognesy\Http\Events\DebugResponseBodyReceived::class,
            Cognesy\Polyglot\Inference\Events\PartialInferenceDeltaCreated::class,
            Cognesy\Polyglot\Inference\Events\StreamEventParsed::class,
        ],
        // Optional class-specific message templates
        'templates' => [],
    ],
];
// @doctest id="5e72"

Logging Presets

Preset	Description
`default`	Development-friendly logging with request enrichment, native-agent and AgentCtrl templates, and noisy streaming events excluded by default
`production`	Minimal logging at `warning` level and above; excludes verbose HTTP, partial-response, and streaming delta events for lower overhead
`custom`	Fully configurable pipeline — supply your own `channel`, `level`, `exclude_events`, `include_events`, and `templates` arrays

Both the default and production presets automatically attach lazy enrichers that add the current HTTP request context (request ID, user ID, session ID, route, method, URL) to every log record.

Events Settings

Configure how Instructor’s internal events are bridged to Laravel’s event dispatcher.

return [
    'events' => [
        // Bridge Instructor events to Laravel's event dispatcher
        'dispatch_to_laravel' => env('INSTRUCTOR_DISPATCH_EVENTS', true),
        // Specific events to bridge (empty = all events)
        'bridge_events' => [
            // \Cognesy\Instructor\Events\ExtractionComplete::class,
        ],
    ],
];
// @doctest id="c8c7"

When bridge_events is empty (the default), every Instructor event is forwarded to Laravel. To limit traffic, list only the event classes you care about. See the Events guide for the full list of available events and listener examples.

Cache Settings

Configure response caching to avoid redundant API calls for identical inputs.

return [
    'cache' => [
        // Enable response caching
        'enabled' => env('INSTRUCTOR_CACHE_ENABLED', false),
        // Cache store to use (null = default store)
        'store' => env('INSTRUCTOR_CACHE_STORE'),
        // Default TTL in seconds
        'ttl' => env('INSTRUCTOR_CACHE_TTL', 3600),
        // Cache key prefix
        'prefix' => 'instructor',
    ],
];
// @doctest id="cc22"

Native Agents Settings

The agents namespace is reserved for native Cognesy\Agents runtime integration. This keeps native runtime, persistence, and observability settings separate from AgentCtrl code-agent execution.

return [
    'agents' => [
        'enabled' => env('INSTRUCTOR_NATIVE_AGENTS_ENABLED', true),
        'session_store' => env('INSTRUCTOR_NATIVE_AGENT_SESSION_STORE', 'memory'),
        'definitions' => [
            base_path('resources/agents'),
        ],
        'tools' => [
            App\Agents\Tools\LookupAccountTool::class,
        ],
        'capabilities' => [
            App\Agents\Capabilities\UseStructuredOutputs::class,
        ],
        'schemas' => [
            'lead' => App\Data\LeadData::class,
        ],
        'broadcasting' => [
            'enabled' => env('INSTRUCTOR_NATIVE_AGENT_BROADCASTING_ENABLED', false),
            'connection' => env('INSTRUCTOR_NATIVE_AGENT_BROADCAST_CONNECTION'),
            'event_name' => env('INSTRUCTOR_NATIVE_AGENT_BROADCAST_EVENT', 'instructor.agent.event'),
            'preset' => env('INSTRUCTOR_NATIVE_AGENT_BROADCAST_PRESET', 'standard'),
        ],
    ],
];
// @doctest id="e1f4"

The Laravel package resolves tools and capabilities through the container before registration, so constructor dependencies are supported. definitions accepts file or directory paths. schemas accepts class strings or schema-definition arrays. If you prefer explicit service-provider wiring, use the first-party tag constants in Cognesy\Instructor\Laravel\Agents\AgentRegistryTags for definitions, tools, capabilities, and tagged SchemaRegistration objects. For long-lived sessions, switch session_store to database and publish the package migration with php artisan vendor:publish --tag=instructor-migrations. For broadcast envelopes, enable agents.broadcasting.enabled and set the Laravel broadcasting connection you want to use.

Code Agents (`AgentCtrl`) Settings

AgentCtrl now reads from the dedicated agent_ctrl namespace. The facade still falls back to the legacy agents key so existing published configs continue to work.

return [
    'agent_ctrl' => [
        'timeout' => env('INSTRUCTOR_AGENT_TIMEOUT', 300),
        'directory' => env('INSTRUCTOR_AGENT_DIRECTORY'),
        'sandbox' => env('INSTRUCTOR_AGENT_SANDBOX', 'host'),
        'claude_code' => [
            'model' => env('CLAUDE_CODE_MODEL', 'claude-sonnet-4-20250514'),
            'timeout' => env('CLAUDE_CODE_TIMEOUT'),
            'directory' => env('CLAUDE_CODE_DIRECTORY'),
            'sandbox' => env('CLAUDE_CODE_SANDBOX'),
        ],
    ],
];
// @doctest id="02c5"

Telemetry Settings

The telemetry namespace is the first-class home for Laravel telemetry wiring.

return [
    'telemetry' => [
        'enabled' => env('INSTRUCTOR_TELEMETRY_ENABLED', false),
        'driver' => env('INSTRUCTOR_TELEMETRY_DRIVER', 'null'),
        'service_name' => env('INSTRUCTOR_TELEMETRY_SERVICE_NAME', env('APP_NAME', 'laravel')),
        'projectors' => [
            'instructor' => true,
            'polyglot' => true,
            'http' => true,
            'agent_ctrl' => true,
            'agents' => true,
        ],
        'http' => [
            'capture_streaming_chunks' => env('INSTRUCTOR_TELEMETRY_HTTP_CAPTURE_STREAMING_CHUNKS', false),
        ],
        'drivers' => [
            'composite' => [
                'exporters' => [],
            ],
            'otel' => [
                'endpoint' => env('INSTRUCTOR_TELEMETRY_OTEL_ENDPOINT'),
                'headers' => [],
            ],
            'langfuse' => [
                'host' => env('INSTRUCTOR_TELEMETRY_LANGFUSE_HOST'),
                'public_key' => env('INSTRUCTOR_TELEMETRY_LANGFUSE_PUBLIC_KEY'),
                'secret_key' => env('INSTRUCTOR_TELEMETRY_LANGFUSE_SECRET_KEY'),
            ],
            'logfire' => [
                'endpoint' => env('INSTRUCTOR_TELEMETRY_LOGFIRE_ENDPOINT'),
                'write_token' => env('INSTRUCTOR_TELEMETRY_LOGFIRE_WRITE_TOKEN'),
                'headers' => [],
            ],
        ],
    ],
];
// @doctest id="9fcc"

driver accepts null, otel, langfuse, logfire, or composite. The package binds telemetry, projector composition, and the runtime event bridge from this namespace automatically.

Environment Variables Reference

Variable	Default	Description
`INSTRUCTOR_CONNECTION`	`openai`	Default LLM connection
`INSTRUCTOR_OUTPUT_MODE`	`json_schema`	Output mode for extraction
`INSTRUCTOR_MAX_RETRIES`	`2`	Max validation retry attempts
`INSTRUCTOR_HTTP_DRIVER`	`laravel`	HTTP client driver
`INSTRUCTOR_HTTP_TIMEOUT`	`120`	Request timeout (seconds)
`INSTRUCTOR_HTTP_CONNECT_TIMEOUT`	`30`	Connection timeout (seconds)
`INSTRUCTOR_LOGGING_ENABLED`	`true`	Enable logging
`INSTRUCTOR_LOG_CHANNEL`	`stack`	Laravel log channel
`INSTRUCTOR_LOG_LEVEL`	`warning`	Minimum log level
`INSTRUCTOR_LOGGING_PRESET`	`production`	Logging preset
`INSTRUCTOR_DISPATCH_EVENTS`	`true`	Bridge events to Laravel
`INSTRUCTOR_CACHE_ENABLED`	`false`	Enable response caching
`OPENAI_API_KEY`	—	OpenAI API key
`ANTHROPIC_API_KEY`	—	Anthropic API key

Runtime Configuration

Override any configuration at runtime using the fluent API on the facades:

use Cognesy\Instructor\Laravel\Facades\StructuredOutput;
use Cognesy\Instructor\StructuredOutputRuntime;

$result = StructuredOutput::connection('anthropic')  // Switch connection
    ->withModel('claude-3-opus-20240229')        // Override model
    ->withRuntime(
        StructuredOutputRuntime::fromDefaults()->withMaxRetries(5)
    )                                            // Override retries
    ->with(
        messages: 'Extract data...',
        responseModel: MyModel::class,
    )
    ->get();
// @doctest id="29e8"

For full programmatic control, build an LLMConfig object and pass it directly:

use Cognesy\Polyglot\Inference\Config\LLMConfig;

$config = LLMConfig::fromArray([
    'driver' => 'openai',
    'apiUrl' => 'https://api.openai.com/v1',
    'apiKey' => $myKey,
    'model' => 'gpt-4o',
    'maxTokens' => 8192,
]);

$result = StructuredOutput::fromConfig($config)
    ->with(messages: '...', responseModel: MyModel::class)
    ->get();
// @doctest id="e527"

Packages

Instructor

Polyglot

Agents

Agent-ctrl

Telemetry

Sandbox

HTTP Client

Laravel

Symfony

Xprompt

Configuration

Configuration

Default Connection

Connections

Supported Drivers

Adding a Custom Connection

Embeddings Connections

Extraction Settings

Output Modes

HTTP Client Settings

Logging Settings

Logging Presets

Events Settings

Cache Settings

Native Agents Settings

Code Agents (`AgentCtrl`) Settings

Telemetry Settings

Environment Variables Reference

Runtime Configuration

Packages

Instructor

Polyglot

Agents

Agent-ctrl

Telemetry

Sandbox

HTTP Client

Laravel

Symfony

Xprompt

​Configuration

​Default Connection

​Connections

​Supported Drivers

​Adding a Custom Connection

​Embeddings Connections

​Extraction Settings

​Output Modes

​HTTP Client Settings

​Logging Settings

​Logging Presets

​Events Settings

​Cache Settings

​Native Agents Settings

​Code Agents (AgentCtrl) Settings

​Telemetry Settings

​Environment Variables Reference

​Runtime Configuration

Configuration

Default Connection

Connections

Supported Drivers

Adding a Custom Connection

Embeddings Connections

Extraction Settings

Output Modes

HTTP Client Settings

Logging Settings

Logging Presets

Events Settings

Cache Settings

Native Agents Settings

Code Agents (`AgentCtrl`) Settings

Telemetry Settings

Environment Variables Reference

Runtime Configuration