schema

type CachedModel struct {
	ServerModel
	Model
	LoadedAt time.Time     `json:"loaded_at,omitzero"`
	Runtime  *ModelRuntime `json:"runtime,omitempty"`
}

type ChatChunk struct {
	Message ChatMessage `json:"message"`
}

type ChatMessage struct {
	Role    string `json:"role"`    // "system", "user", "assistant", or "tool"
	Content string `json:"content"` // The message content
}

type ChatRequest struct {
	CompletionRequest
	Messages []ChatMessage `json:"messages"`
}

type ChatResponse struct {
	Model        string       `json:"model"`                   // Model used
	Thinking     *ChatMessage `json:"thinking,omitempty"`      // Optional reasoning message
	Message      ChatMessage  `json:"message"`                 // Assistant message
	Usage        Usage        `json:"usage"`                   // Token usage
	FinishReason string       `json:"finish_reason,omitempty"` // Reason generation ended
}

type CompletionChunk struct {
	Text string `json:"text"` // Chunk text
}

type CompletionRequest struct {
	Model         string   `json:"model"`                    // Model name
	Prompt        string   `json:"prompt"`                   // Prompt to complete
	MaxTokens     *int32   `json:"max_tokens,omitempty"`     // Max tokens to generate
	Temperature   *float32 `json:"temperature,omitempty"`    // Sampling temperature
	TopP          *float32 `json:"top_p,omitempty"`          // Nucleus sampling
	TopK          *int32   `json:"top_k,omitempty"`          // Top-k sampling
	RepeatPenalty *float32 `json:"repeat_penalty,omitempty"` // Penalize repeats (1.0 = disabled)
	RepeatLastN   *int32   `json:"repeat_last_n,omitempty"`  // Repeat penalty window size
	Seed          *uint32  `json:"seed,omitempty"`           // RNG seed
	Stop          []string `json:"stop,omitempty"`           // Stop words
	PrefixCache   *bool    `json:"prefix_cache,omitempty"`   // Enable prefix caching
}

type CompletionResponse struct {
	Model        string `json:"model"`                   // Model used
	Text         string `json:"text"`                    // Completion text
	Usage        Usage  `json:"usage"`                   // Token usage
	FinishReason string `json:"finish_reason,omitempty"` // Reason generation ended
}

type ContextRequest struct {
	LoadModelRequest
	ContextSize   *uint32 `json:"context_size,omitempty"`   // Context size (nil = from model)
	BatchSize     *uint32 `json:"batch_size,omitempty"`     // Logical batch size (nil = default)
	UBatchSize    *uint32 `json:"ubatch_size,omitempty"`    // Physical/micro batch size (nil = default, must equal batch_size for encoder models)
	Threads       *int32  `json:"threads,omitempty"`        // Number of threads (nil = default)
	AttentionType *int32  `json:"attention_type,omitempty"` // Attention type: -1=auto, 0=causal, 1=non-causal (nil = auto)
	FlashAttn     *int32  `json:"flash_attn,omitempty"`     // Flash attention: -1=auto, 0=disabled, 1=enabled (nil = auto)
	Embeddings    *bool   `json:"embeddings,omitempty"`     // Enable embeddings extraction (nil = false)
	KVUnified     *bool   `json:"kv_unified,omitempty"`     // Use unified KV cache (nil = default, required for BERT)
}

type DetokenizeRequest struct {
	Model          string  `json:"model"`                     // Model name or path (must be loaded)
	Tokens         []Token `json:"tokens"`                    // Tokens to detokenize
	RemoveSpecial  *bool   `json:"remove_special,omitempty"`  // Remove BOS/EOS tokens (default: false)
	UnparseSpecial *bool   `json:"unparse_special,omitempty"` // Render special tokens as text (default: true)
}

type DetokenizeResponse struct {
	Text string `json:"text"`
}

type EmbedRequest struct {
	Model     string   `json:"model"`               // Model name
	Input     []string `json:"input"`               // Text(s) to embed
	Normalize *bool    `json:"normalize,omitempty"` // L2-normalize embeddings (default: true)
}

type EmbedResponse struct {
	Model      string      `json:"model"`      // Model used
	Embeddings [][]float32 `json:"embeddings"` // One embedding vector per input
	Dimension  int         `json:"dimension"`  // Embedding dimension
	Usage      Usage       `json:"usage"`      // Token usage
}

type GPUDevice struct {
	ID               int32  `json:"id"`
	Name             string `json:"name"`
	FreeMemoryBytes  int64  `json:"free_memory_bytes"`  // -1 if unknown
	TotalMemoryBytes int64  `json:"total_memory_bytes"` // -1 if unknown
}

type GPUInfo struct {
	Backend string      `json:"backend"` // "Metal", "CUDA", "Vulkan", "CPU"
	Devices []GPUDevice `json:"devices"`
}

type LoadModelRequest struct {
	Name   string `json:"name"`                 // Model name or path to load
	Load   *bool  `json:"load,omitempty"`       // Load (true) or unload (false) model (nil = load)
	Gpu    *int32 `json:"gpu,omitempty"`        // Main GPU index (nil = default)
	Layers *int32 `json:"gpu_layers,omitempty"` // Number of layers to offload to GPU (nil = default, -1 = all)
	Mmap   *bool  `json:"use_mmap,omitempty"`   // Use memory mapping for model loading (nil = default)
	Mlock  *bool  `json:"use_mlock,omitempty"`  // Lock model in memory (nil = default)
}

type Model struct {
	// Identity
	Path         string `json:"path,omitempty"`
	Name         string `json:"name,omitempty"`
	Architecture string `json:"architecture,omitempty"`
	Description  string `json:"description,omitempty"`

	// Chat template
	ChatTemplate string `json:"chatTemplate,omitempty"`

	// Dimensions
	ContextSize   int32 `json:"contextSize,omitempty"`
	EmbeddingSize int32 `json:"embeddingSize,omitempty"`
	LayerCount    int32 `json:"layerCount,omitempty"`
	HeadCount     int32 `json:"headCount,omitempty"`
	HeadKVCount   int32 `json:"headKVCount,omitempty"`

	// Raw metadata key/value pairs from the model
	Meta map[string]any `json:"meta,omitempty"`
}

type ModelPullProgress struct {
	Filename      string  `json:"model"`
	BytesReceived uint64  `json:"bytes_received"`
	TotalBytes    uint64  `json:"total_bytes,omitempty"`
	Percentage    float64 `json:"percent,omitempty"`
}

type ModelRuntime struct {
	NLayer    int32  `json:"layerCount,omitempty"`
	NHead     int32  `json:"headCount,omitempty"`
	NHeadKV   int32  `json:"headKVCount,omitempty"`
	NEmbd     int32  `json:"embeddingSize,omitempty"`
	NCtxTrain int32  `json:"contextSize,omitempty"`
	NParams   uint64 `json:"paramCount,omitempty"`
	ModelSize uint64 `json:"modelSizeBytes,omitempty"`
}

type PullModelRequest struct {
	URL string `json:"url"` // URL to download the model from (supports hf:// and https://)
}

type ServerModel struct {
	sync.RWMutex
	Handle *llamacpp.Model
}

type Token = int32

type TokenizeRequest struct {
	Model        string `json:"model"`                   // Model name or path (must be loaded)
	Text         string `json:"text"`                    // Text to tokenize
	AddSpecial   *bool  `json:"add_special,omitempty"`   // Add BOS/EOS tokens (default: true)
	ParseSpecial *bool  `json:"parse_special,omitempty"` // Parse special tokens in text (default: false)
}

type TokenizeResponse struct {
	Tokens []Token `json:"tokens"`
}

type Usage struct {
	InputTokens  int `json:"input_tokens"`  // Tokens in input (prompt/text to embed)
	OutputTokens int `json:"output_tokens"` // Tokens generated (0 for embeddings)
}