config

type ByteSize uint64

type Config struct {
	LLM        LLM        `toml:"llm"`
	Documents  Documents  `toml:"documents"`
	Extraction Extraction `toml:"extraction"`

	// Warnings collects non-fatal messages (e.g. deprecations) during load.
	// Not serialized; the caller decides how to display them.
	Warnings []string `toml:"-"`
}

type Documents struct {
	// MaxFileSize is the largest file that can be imported as a document
	// attachment. Accepts unitized strings ("50 MiB") or bare integers
	// (bytes). Default: 50 MiB.
	MaxFileSize ByteSize `toml:"max_file_size" env:"MICASA_MAX_DOCUMENT_SIZE"`

	// CacheTTL is the preferred cache lifetime setting. Accepts unitized
	// strings ("30d", "720h") or bare integers (seconds). Default: 30d.
	CacheTTL *Duration `toml:"cache_ttl,omitempty" env:"MICASA_CACHE_TTL"`

	// CacheTTLDays is deprecated; use CacheTTL instead. Kept for backward
	// compatibility. Bare integer interpreted as days.
	CacheTTLDays *int `toml:"cache_ttl_days,omitempty" env:"MICASA_CACHE_TTL_DAYS"`
}

type Duration struct{ time.Duration }

type Extraction struct {
	// Model overrides llm.model for extraction. Extraction wants a small,
	// fast model optimized for structured JSON output. Defaults to the
	// chat model if empty.
	Model string `toml:"model" env:"MICASA_EXTRACTION_MODEL"`

	// MaxExtractPages is the maximum number of pages for async extraction of scanned
	// documents. Front-loaded info (specs, warranty) is typically in the
	// first pages. Default: 20.
	MaxExtractPages int `toml:"max_extract_pages" env:"MICASA_MAX_EXTRACT_PAGES"`

	// Enabled controls whether LLM-powered extraction runs when a document
	// is uploaded. Text and image extraction are independent and always
	// available. Default: true.
	Enabled *bool `toml:"enabled,omitempty" env:"MICASA_EXTRACTION_ENABLED"`

	// TextTimeout is the maximum time to wait for pdftotext. Go duration
	// string, e.g. "30s", "1m". Default: "30s".
	TextTimeout string `toml:"text_timeout" env:"MICASA_TEXT_TIMEOUT"`

	// Thinking enables the model's internal reasoning mode (e.g. qwen3
	// <think> blocks). Disable for faster responses when structured output
	// is all you need. Default: false.
	Thinking *bool `toml:"thinking,omitempty" env:"MICASA_EXTRACTION_THINKING"`
}

type LLM struct {
	// BaseURL is the root of an OpenAI-compatible API.
	// The client appends /chat/completions, /models, etc.
	// Default: http://localhost:11434/v1 (Ollama)
	BaseURL string `toml:"base_url" env:"OLLAMA_HOST"`

	// Model is the model identifier passed in chat requests.
	// Default: qwen3
	Model string `toml:"model" env:"MICASA_LLM_MODEL"`

	// ExtraContext is custom text appended to all system prompts.
	// Useful for domain-specific details: house style, currency, location, etc.
	// Optional; defaults to empty.
	ExtraContext string `toml:"extra_context"`

	// Timeout is the maximum time to wait for quick LLM server operations
	// (ping, model listing, auto-detect). Go duration string, e.g. "5s",
	// "10s", "500ms". Default: "5s".
	Timeout string `toml:"timeout" env:"MICASA_LLM_TIMEOUT"`

	// Thinking enables the model's internal reasoning mode for chat
	// (e.g. qwen3 <think> blocks). Default: unset (not sent to server).
	Thinking *bool `toml:"thinking,omitempty" env:"MICASA_LLM_THINKING"`
}