---
title: Configuration
description: Field-level configuration behavior for the AI extension
deprecated: true
supersededBy: ['/docs/cortex', '/docs/shield', '/docs/sentinel', '/docs/weave', '/docs/nexus']
---

## Config Model

The extension config is `ai.Config` and is passed programmatically (`ai.NewExtension(...)` or `ai.NewExtensionWithConfig(...)`).

## Top-Level Fields (`ai.Config`)

- `EnableLLM` (default `true`)
- `EnableAgents` (default `true`)
- `EnableTraining` (default `false`)
- `EnableInference` (default `true`)
- `EnableCoordination` (default `true`)
- `MaxConcurrency` (default `10`)
- `RequestTimeout` (default `30s`)
- `CacheSize` (default `1000`)
- `LLM` (`LLMConfiguration`)
- `Inference` (`InferenceConfiguration`)
- `Agents` (`AgentConfiguration`)
- `Middleware` (`MiddlewareConfiguration`)
- `StateStore` (`StateStoreConfig`)
- `VectorStore` (`VectorStoreConfig`)
- `Training` (`TrainingConfiguration`)

## LLM Configuration (`LLMConfiguration`)

Fields:

- `DefaultProvider`
- `Providers`
- `MaxRetries`
- `RetryDelay`
- `Timeout`

Important behavior:

- `CreateLLMManager` uses `DefaultProvider`, retries, and timeout.
- Providers from `Providers` are currently **not** auto-registered.
- Register providers manually before startup for real model calls.

## State Store (`StateStoreConfig`)

`Type` values:

- `memory` (default)
- `postgres`
- `redis`

Sub-configs:

- `Memory.TTL`
- `Postgres.ConnString`, `Postgres.TableName`
- `Redis.Addrs`, `Redis.Password`, `Redis.DB`

Behavior:

- invalid config or init errors trigger memory fallback in extension registration.

## Vector Store (`VectorStoreConfig`)

`Type` values:

- `memory` (supported)
- `postgres` (declared, currently not implemented)
- `pinecone` (declared, currently not implemented)
- `weaviate` (declared, currently not implemented)

Behavior:

- unsupported/non-implemented backends return errors in factory code.
- extension then falls back to in-memory vector store.

## Training (`TrainingConfiguration`)

Important fields:

- `Enabled`
- `CheckpointPath`
- `ModelPath`
- `DataPath`
- `MaxConcurrentJobs`
- `DefaultResources` (`cpu`, `memory`, `gpu`, `timeout`, `priority`)
- `Storage` (`local`, `s3`, `gcs`, `azure`)

Important behavior:

- DI registration of training services is controlled by `Training.Enabled`.

## Inference, Agents, Middleware Blocks

`Inference`, `Agents`, and `Middleware` config structs are available for package-level modules and future extension wiring.

Current extension behavior does not automatically instantiate inference engine or middleware from these blocks.

## Practical Production Config (Go)

```go
cfg := ai.DefaultConfig()

cfg.LLM.DefaultProvider = "openai"
cfg.LLM.MaxRetries = 3
cfg.LLM.Timeout = 30 * time.Second

cfg.StateStore = ai.StateStoreConfig{
    Type: "postgres",
    Postgres: &ai.PostgresStateConfig{
        ConnString: os.Getenv("DATABASE_URL"),
        TableName:  "agent_state",
    },
}

// Keep memory unless you have a custom vector constructor.
cfg.VectorStore = ai.VectorStoreConfig{Type: "memory"}

cfg.Training = ai.TrainingConfiguration{
    Enabled:           true,
    CheckpointPath:    "./checkpoints",
    ModelPath:         "./models",
    DataPath:          "./data",
    MaxConcurrentJobs: 4,
    DefaultResources: ai.ResourcesConfig{
        CPU:      "4",
        Memory:   "8Gi",
        GPU:      0,
        Timeout:  4 * time.Hour,
        Priority: 1,
    },
    Storage: ai.StorageConfig{
        Type: "local",
        Local: &ai.LocalStorageConfig{BasePath: "./training-storage"},
    },
}

ext := ai.NewExtension(ai.WithConfig(cfg))
```

## Validation Tips

- Fail fast on startup by keeping constructors eager.
- Verify your provider can execute a simple prompt after app start.
- Confirm state and vector store effective types from startup logs.