@yourgpt/llm-sdk 2.0.1 → 2.0.2-beta.1

package/dist/{base-DdxolpKP.d.ts → types-ClJBhmS0.d.ts}

@@ -342,7 +342,7 @@ interface ResponseOptions {
 /**
  * Stream event types
  */
-type StreamEventType = "message:start" | "message:delta" | "message:end" | "thinking:start" | "thinking:delta" | "thinking:end" | "action:start" | "action:args" | "action:end" | "tool_calls" | "tool:result" | "loop:iteration" | "loop:complete" | "error" | "done";
+type StreamEventType = "message:start" | "message:delta" | "message:end" | "thinking:start" | "thinking:delta" | "thinking:end" | "action:start" | "action:args" | "action:end" | "tool_calls" | "tool:result" | "citation" | "loop:iteration" | "loop:complete" | "error" | "done";
 /**
  * Base event interface
  */
@@ -479,6 +479,30 @@ interface LoopCompleteEvent extends BaseEvent {
     aborted?: boolean;
     maxIterationsReached?: boolean;
 }
+/**
+ * Citation from web search (unified format for all providers)
+ */
+interface Citation {
+    /** Unique citation index (1-based) */
+    index: number;
+    /** Source URL */
+    url: string;
+    /** Page title */
+    title: string;
+    /** Cited text snippet (optional) */
+    citedText?: string;
+    /** Source domain (extracted from URL) */
+    domain?: string;
+    /** Favicon URL (generated from domain) */
+    favicon?: string;
+}
+/**
+ * Citation event - web search returned citations
+ */
+interface CitationEvent extends BaseEvent {
+    type: "citation";
+    citations: Citation[];
+}
 /**
  * Message format for done event (API format with snake_case)
  */
@@ -516,7 +540,7 @@ interface DoneEvent extends BaseEvent {
 /**
  * Union of all stream events
  */
-type StreamEvent = MessageStartEvent | MessageDeltaEvent | MessageEndEvent | ThinkingStartEvent | ThinkingDeltaEvent | ThinkingEndEvent | ActionStartEvent | ActionArgsEvent | ActionEndEvent | ToolCallsEvent | ToolResultEvent | LoopIterationEvent | LoopCompleteEvent | ErrorEvent | DoneEvent;
+type StreamEvent = MessageStartEvent | MessageDeltaEvent | MessageEndEvent | ThinkingStartEvent | ThinkingDeltaEvent | ThinkingEndEvent | ActionStartEvent | ActionArgsEvent | ActionEndEvent | ToolCallsEvent | ToolResultEvent | CitationEvent | LoopIterationEvent | LoopCompleteEvent | ErrorEvent | DoneEvent;
 /**
  * LLM configuration
  */
@@ -671,6 +695,49 @@ interface AgentLoopConfig {
     debug?: boolean;
     enabled?: boolean;
 }
+/**
+ * Web search configuration for native provider search
+ *
+ * Enables native web search for supported providers:
+ * - Anthropic: Uses Claude's built-in web search tool
+ * - OpenAI: Uses GPT's web search preview
+ * - Google: Uses Gemini's Google Search grounding
+ *
+ * @example
+ * ```typescript
+ * const runtime = createRuntime({
+ *   provider: createAnthropic({ apiKey: '...' }),
+ *   model: 'claude-sonnet-4-20250514',
+ *   webSearch: true, // Enable with defaults
+ * });
+ *
+ * // Or with configuration
+ * const runtime = createRuntime({
+ *   provider: createOpenAI({ apiKey: '...' }),
+ *   model: 'gpt-4o',
+ *   webSearch: {
+ *     maxUses: 5,
+ *     allowedDomains: ['docs.anthropic.com', 'openai.com'],
+ *   },
+ * });
+ * ```
+ */
+interface WebSearchConfig {
+    /** Maximum number of search uses per request (default: unlimited) */
+    maxUses?: number;
+    /** Only search these domains (provider-specific support) */
+    allowedDomains?: string[];
+    /** Exclude these domains from search (provider-specific support) */
+    blockedDomains?: string[];
+    /** User location for localized results (Anthropic only) */
+    userLocation?: {
+        type: "approximate";
+        city?: string;
+        region?: string;
+        country?: string;
+        timezone?: string;
+    };
+}
 /**
  * Unified tool call format
  */
@@ -741,6 +808,11 @@ interface ChatCompletionRequest {
     config?: RequestLLMConfig;
     /** Abort signal for cancellation */
     signal?: AbortSignal;
+    /**
+     * Enable native web search for the provider.
+     * When true or configured, the provider's native search is enabled.
+     */
+    webSearch?: boolean | WebSearchConfig;
 }
 /**
  * Non-streaming completion result
@@ -955,4 +1027,356 @@ type OpenAIMessage = {
  */
 declare function formatMessagesForOpenAI(messages: Message[], systemPrompt?: string): OpenAIMessage[];
 
-export { formatTools as $, type ActionDefinition as A, type FinishChunk as B, type CoreMessage as C, type DoneEventMessage as D, type ErrorChunk as E, type FilePart as F, type GenerateTextParams as G, type TokenUsage as H, type ImagePart as I, type FinishReason as J, type KnowledgeBaseConfig as K, type LLMAdapter as L, type Message as M, DEFAULT_CAPABILITIES as N, type ChatCompletionRequest as O, type AdapterFactory as P, type LLMConfig as Q, type ResponseOptions as R, type StreamTextParams as S, type ToolContext$1 as T, type UserMessage as U, type ToolLocation as V, type UnifiedToolCall as W, type UnifiedToolResult as X, type ToolExecution as Y, type CompletionResult as Z, formatMessages as _, type GenerateTextResult as a, formatMessagesForAnthropic as a0, formatMessagesForOpenAI as a1, messageToAnthropicContent as a2, messageToOpenAIContent as a3, hasImageAttachments as a4, hasMediaAttachments as a5, attachmentToAnthropicImage as a6, attachmentToAnthropicDocument as a7, attachmentToOpenAIImage as a8, type AnthropicContentBlock as a9, type OpenAIContentBlock as aa, type StreamTextResult as b, type Tool as c, type ToolDefinition as d, type AgentLoopConfig as e, type StreamEvent as f, type ToolCallInfo as g, type TokenUsageRaw as h, type ToolResponse as i, type LanguageModel as j, type ModelCapabilities as k, type DoGenerateParams as l, type DoGenerateResult as m, type SystemMessage as n, type AssistantMessage as o, type ToolMessage as p, type UserContentPart as q, type TextPart as r, type ToolCall$1 as s, type ToolResult as t, type GenerateStep as u, type StreamPart as v, type StreamChunk as w, type TextDeltaChunk as x, type ToolCallChunk as y, type ToolResultChunk as z };
+/**
+ * Provider Types
+ *
+ * Defines interfaces for:
+ * 1. Provider Formatters (for tool transformations in agent loop)
+ * 2. Multi-provider architecture (AIProvider, capabilities, configs)
+ */
+
+/**
+ * Provider formatter interface
+ *
+ * Each provider implements this interface to handle:
+ * - Tool definition transformation
+ * - Tool call parsing from responses
+ * - Tool result formatting
+ * - Stop reason detection
+ */
+interface ProviderFormatter {
+    /**
+     * Transform unified tool definitions to provider format
+     */
+    transformTools(tools: ToolDefinition[]): unknown[];
+    /**
+     * Parse tool calls from provider response
+     */
+    parseToolCalls(response: unknown): UnifiedToolCall[];
+    /**
+     * Format tool results for provider
+     */
+    formatToolResults(results: UnifiedToolResult[]): unknown[];
+    /**
+     * Check if response indicates tool use is requested
+     */
+    isToolUseStop(response: unknown): boolean;
+    /**
+     * Check if response indicates end of turn
+     */
+    isEndTurnStop(response: unknown): boolean;
+    /**
+     * Get stop reason string from response
+     */
+    getStopReason(response: unknown): string;
+    /**
+     * Extract text content from response
+     */
+    extractTextContent(response: unknown): string;
+    /**
+     * Build assistant message with tool calls for conversation history
+     */
+    buildAssistantToolMessage(toolCalls: UnifiedToolCall[], textContent?: string): unknown;
+    /**
+     * Build user message with tool results for conversation history
+     */
+    buildToolResultMessage(results: UnifiedToolResult[]): unknown;
+}
+/**
+ * Anthropic tool definition format
+ */
+interface AnthropicTool {
+    name: string;
+    description: string;
+    input_schema: {
+        type: "object";
+        properties: Record<string, unknown>;
+        required?: string[];
+    };
+}
+/**
+ * Anthropic tool_use block from response
+ */
+interface AnthropicToolUse {
+    type: "tool_use";
+    id: string;
+    name: string;
+    input: Record<string, unknown>;
+}
+/**
+ * Anthropic tool_result block
+ */
+interface AnthropicToolResult {
+    type: "tool_result";
+    tool_use_id: string;
+    content: string;
+}
+/**
+ * OpenAI tool definition format
+ */
+interface OpenAITool {
+    type: "function";
+    function: {
+        name: string;
+        description: string;
+        parameters: {
+            type: "object";
+            properties: Record<string, unknown>;
+            required?: string[];
+        };
+    };
+}
+/**
+ * OpenAI tool call from response
+ */
+interface OpenAIToolCall {
+    id: string;
+    type: "function";
+    function: {
+        name: string;
+        arguments: string;
+    };
+}
+/**
+ * OpenAI tool result message
+ */
+interface OpenAIToolResult {
+    role: "tool";
+    tool_call_id: string;
+    content: string;
+}
+/**
+ * Google Gemini function declaration
+ */
+interface GeminiFunctionDeclaration {
+    name: string;
+    description: string;
+    parameters?: {
+        type: "object";
+        properties: Record<string, unknown>;
+        required?: string[];
+    };
+}
+/**
+ * Gemini function call from response
+ */
+interface GeminiFunctionCall {
+    name: string;
+    args: Record<string, unknown>;
+}
+/**
+ * Gemini function response
+ */
+interface GeminiFunctionResponse {
+    name: string;
+    response: Record<string, unknown>;
+}
+/**
+ * Capabilities of a model for UI feature flags
+ * UI components can use this to enable/disable features
+ */
+interface ProviderCapabilities {
+    /** Supports image inputs */
+    supportsVision: boolean;
+    /** Supports tool/function calling */
+    supportsTools: boolean;
+    /** Supports extended thinking (Claude, DeepSeek) */
+    supportsThinking: boolean;
+    /** Supports streaming responses */
+    supportsStreaming: boolean;
+    /** Supports PDF document inputs */
+    supportsPDF: boolean;
+    /** Supports audio inputs */
+    supportsAudio: boolean;
+    /** Supports video inputs */
+    supportsVideo: boolean;
+    /** Maximum context tokens */
+    maxTokens: number;
+    /** Supported image MIME types */
+    supportedImageTypes: string[];
+    /** Supported audio MIME types */
+    supportedAudioTypes?: string[];
+    /** Supported video MIME types */
+    supportedVideoTypes?: string[];
+    /** Supports JSON mode / structured output */
+    supportsJsonMode?: boolean;
+    /** Supports system messages */
+    supportsSystemMessages?: boolean;
+}
+/**
+ * AI Provider interface (object form)
+ *
+ * Wraps existing LLMAdapter with additional metadata:
+ * - Supported models list
+ * - Per-model capabilities
+ * - Provider name
+ */
+interface AIProviderObject {
+    /** Provider name (e.g., 'openai', 'anthropic') */
+    readonly name: string;
+    /** List of supported model IDs */
+    readonly supportedModels: string[];
+    /**
+     * Get a language model adapter for the given model ID
+     * Returns the existing LLMAdapter interface - no breaking changes
+     */
+    languageModel(modelId: string): LLMAdapter;
+    /**
+     * Get capabilities for a specific model
+     * UI components use this to enable/disable features
+     */
+    getCapabilities(modelId: string): ProviderCapabilities;
+    /**
+     * Optional: Get an embedding model (future expansion)
+     */
+    embeddingModel?(modelId: string): EmbeddingModel;
+}
+/**
+ * Callable AI Provider (Vercel AI SDK style)
+ *
+ * A function that returns a LanguageModel when called with a model ID,
+ * but also has properties for provider metadata and methods.
+ *
+ * @example
+ * ```typescript
+ * const openai = createOpenAI({ apiKey: '...' });
+ *
+ * // Callable - returns LanguageModel directly (Vercel AI SDK style)
+ * const model = openai('gpt-4o');
+ *
+ * // Also supports method calls (backward compatible)
+ * const model2 = openai.languageModel('gpt-4o');
+ *
+ * // Check capabilities
+ * const caps = openai.getCapabilities('gpt-4o');
+ * if (caps.supportsVision) {
+ *   // Show image upload button
+ * }
+ * ```
+ */
+interface AIProvider extends AIProviderObject {
+    /**
+     * Call the provider directly with a model ID to get a LanguageModel
+     * This is the Vercel AI SDK style pattern
+     */
+    (modelId: string): LLMAdapter;
+}
+/**
+ * Embedding model interface (for future expansion)
+ */
+interface EmbeddingModel {
+    readonly provider: string;
+    readonly modelId: string;
+    embed(texts: string[]): Promise<number[][]>;
+}
+/**
+ * Base provider configuration
+ */
+interface BaseProviderConfig {
+    /** API key (falls back to environment variable) */
+    apiKey?: string;
+    /** Custom base URL */
+    baseUrl?: string;
+    /** Request timeout in milliseconds */
+    timeout?: number;
+    /** Custom headers to include */
+    headers?: Record<string, string>;
+}
+/**
+ * OpenAI provider configuration
+ */
+interface OpenAIProviderConfig extends BaseProviderConfig {
+    /** OpenAI organization ID */
+    organization?: string;
+    /** OpenAI project ID */
+    project?: string;
+    /** Vision detail level for images */
+    imageDetail?: "auto" | "low" | "high";
+}
+/**
+ * Anthropic provider configuration
+ */
+interface AnthropicProviderConfig extends BaseProviderConfig {
+    /** Extended thinking budget in tokens (minimum 1024) */
+    thinkingBudget?: number;
+    /** Enable prompt caching */
+    cacheControl?: boolean;
+}
+/**
+ * Google provider configuration
+ */
+interface GoogleProviderConfig extends BaseProviderConfig {
+    /** Safety settings */
+    safetySettings?: GoogleSafetySetting[];
+    /** Grounding configuration (for web search) */
+    groundingConfig?: GoogleGroundingConfig;
+}
+/**
+ * Google safety setting
+ */
+interface GoogleSafetySetting {
+    category: "HARM_CATEGORY_HARASSMENT" | "HARM_CATEGORY_HATE_SPEECH" | "HARM_CATEGORY_SEXUALLY_EXPLICIT" | "HARM_CATEGORY_DANGEROUS_CONTENT";
+    threshold: "BLOCK_NONE" | "BLOCK_LOW_AND_ABOVE" | "BLOCK_MEDIUM_AND_ABOVE" | "BLOCK_HIGH_AND_ABOVE";
+}
+/**
+ * Google grounding configuration
+ */
+interface GoogleGroundingConfig {
+    /** Enable Google Search grounding */
+    googleSearchRetrieval?: boolean;
+}
+/**
+ * xAI provider configuration
+ */
+interface XAIProviderConfig extends BaseProviderConfig {
+}
+/**
+ * Azure OpenAI provider configuration
+ */
+interface AzureProviderConfig extends BaseProviderConfig {
+    /** Azure resource name */
+    resourceName: string;
+    /** Deployment name */
+    deploymentName: string;
+    /** API version (default: 2024-02-15-preview) */
+    apiVersion?: string;
+}
+/**
+ * Ollama model-specific options
+ * These map to Ollama's native API options
+ */
+interface OllamaModelOptions {
+    /** Context window size (default varies by model) */
+    num_ctx?: number;
+    /** Max tokens to predict (-1 = infinite, -2 = fill context) */
+    num_predict?: number;
+    /** Mirostat sampling (0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) */
+    mirostat?: 0 | 1 | 2;
+    /** Mirostat learning rate (default: 0.1) */
+    mirostat_eta?: number;
+    /** Mirostat target entropy (default: 5.0) */
+    mirostat_tau?: number;
+    /** Repeat penalty (default: 1.1) */
+    repeat_penalty?: number;
+    /** Random seed for reproducibility (-1 = random) */
+    seed?: number;
+    /** Top-k sampling (default: 40) */
+    top_k?: number;
+    /** Top-p (nucleus) sampling (default: 0.9) */
+    top_p?: number;
+    /** Min-p sampling (default: 0.0) */
+    min_p?: number;
+    /** Stop sequences */
+    stop?: string[];
+    /** Temperature override (also available in config) */
+    temperature?: number;
+}
+/**
+ * Ollama provider configuration
+ */
+interface OllamaProviderConfig extends BaseProviderConfig {
+    /** Default Ollama-specific model options */
+    options?: OllamaModelOptions;
+}
+
+export { type XAIProviderConfig as $, type AIProvider as A, type ToolResultChunk as B, type CoreMessage as C, type DoneEventMessage as D, type FinishChunk as E, type FilePart as F, type GenerateTextParams as G, type ErrorChunk as H, type ImagePart as I, type TokenUsage as J, type KnowledgeBaseConfig as K, type LLMAdapter as L, type Message as M, type FinishReason as N, DEFAULT_CAPABILITIES as O, type ChatCompletionRequest as P, type AdapterFactory as Q, type ResponseOptions as R, type StreamTextParams as S, type ToolContext$1 as T, type UserMessage as U, type ProviderCapabilities as V, type WebSearchConfig as W, type BaseProviderConfig as X, type OpenAIProviderConfig as Y, type AnthropicProviderConfig as Z, type GoogleProviderConfig as _, type GenerateTextResult as a, type AzureProviderConfig as a0, type OllamaProviderConfig as a1, type OllamaModelOptions as a2, type ProviderFormatter as a3, type AnthropicTool as a4, type AnthropicToolUse as a5, type AnthropicToolResult as a6, type OpenAITool as a7, type OpenAIToolCall as a8, type OpenAIToolResult as a9, type GeminiFunctionDeclaration as aa, type GeminiFunctionCall as ab, type GeminiFunctionResponse as ac, type LLMConfig as ad, type ToolLocation as ae, type UnifiedToolCall as af, type UnifiedToolResult as ag, type ToolExecution as ah, type Citation as ai, type CompletionResult as aj, formatMessages as ak, formatTools as al, formatMessagesForAnthropic as am, formatMessagesForOpenAI as an, messageToAnthropicContent as ao, messageToOpenAIContent as ap, hasImageAttachments as aq, hasMediaAttachments as ar, attachmentToAnthropicImage as as, attachmentToAnthropicDocument as at, attachmentToOpenAIImage as au, type AnthropicContentBlock as av, type OpenAIContentBlock as aw, type StreamTextResult as b, type Tool as c, type ActionDefinition as d, type ToolDefinition as e, type AgentLoopConfig as f, type StreamEvent as g, type ToolCallInfo as h, type TokenUsageRaw as i, type ToolResponse as j, type LanguageModel as k, type ModelCapabilities as l, type DoGenerateParams as m, type DoGenerateResult as n, type SystemMessage as o, type AssistantMessage as p, type ToolMessage as q, type UserContentPart as r, type TextPart as s, type ToolCall$1 as t, type ToolResult as u, type GenerateStep as v, type StreamPart as w, type StreamChunk as x, type TextDeltaChunk as y, type ToolCallChunk as z };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yourgpt/llm-sdk",
-  "version": "2.0.1",
+  "version": "2.0.2-beta.1",
   "description": "AI SDK for building AI Agents with any LLM",
   "main": "./dist/index.js",
   "module": "./dist/index.mjs",