@illuma-ai/agents 1.0.81

package/dist/cjs/llm/bedrock/index.cjs

@@ -0,0 +1,282 @@
+'use strict';
+
+var aws = require('@langchain/aws');
+var messages = require('@langchain/core/messages');
+var outputs = require('@langchain/core/outputs');
+
+/**
+ * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts
+ * and adds support for:
+ *
+ * - Prompt caching support for Bedrock Converse API (Illuma feature)
+ * - Application Inference Profiles (PR #9129)
+ * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+
+ *
+ * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,
+ * causing LangChain's merge logic to fail with "field[contentBlockIndex] already exists"
+ * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid
+ * the conflict.
+ *
+ * The contentBlockIndex field is only used internally by Bedrock's streaming protocol
+ * and isn't needed by application logic - the index field on tool_call_chunks serves
+ * the purpose of tracking tool call ordering.
+ *
+ * PROMPT CACHING:
+ * When promptCache: true is set, this wrapper adds cachePoint markers to the tools array
+ * to enable Bedrock prompt caching for tool definitions. This allows tool schemas to be
+ * cached and reused across requests, reducing latency and costs.
+ *
+ * CACHE TOKEN EXTRACTION:
+ * LangChain AWS doesn't extract cacheReadInputTokens/cacheWriteInputTokens from Bedrock's
+ * response. This wrapper adds input_token_details to usage_metadata with cache information.
+ */
+class CustomChatBedrockConverse extends aws.ChatBedrockConverse {
+    /** Enable Bedrock prompt caching for tool definitions */
+    promptCache;
+    /** Application Inference Profile ARN to use instead of model ID */
+    applicationInferenceProfile;
+    /** Service tier for model invocation */
+    serviceTier;
+    constructor(fields) {
+        super(fields);
+        this.promptCache = fields?.promptCache ?? false;
+        this.applicationInferenceProfile = fields?.applicationInferenceProfile;
+        this.serviceTier = fields?.serviceTier;
+        // Fix: Force supportsToolChoiceValues for Claude models
+        // The parent constructor checks `model.includes('claude-3')` but this fails when:
+        // 1. Using applicationInferenceProfile ARNs (arn:aws:bedrock:...)
+        // 2. Using different naming conventions (claude-4, claude-opus-4, etc.)
+        // We need to ensure tool_choice is properly set for withStructuredOutput to work
+        const modelName = (fields?.model ?? '').toLowerCase();
+        const profileName = (fields?.applicationInferenceProfile ?? '').toLowerCase();
+        const isClaudeModel = modelName.includes('claude') ||
+            modelName.includes('anthropic') ||
+            profileName.includes('claude') ||
+            profileName.includes('anthropic');
+        if (isClaudeModel && !this.supportsToolChoiceValues?.length) {
+            // Claude models support all tool choice values
+            this.supportsToolChoiceValues = ['auto', 'any', 'tool'];
+        }
+    }
+    static lc_name() {
+        return 'IllumaBedrockConverse';
+    }
+    /**
+     * Get the model ID to use for API calls.
+     * Returns applicationInferenceProfile if set, otherwise returns this.model.
+     */
+    getModelId() {
+        return this.applicationInferenceProfile ?? this.model;
+    }
+    /**
+     * Override invocationParams to:
+     * 1. Add cachePoint to tools when promptCache is enabled
+     * 2. Add serviceTier support
+     *
+     * CACHING STRATEGY: Separate cachePoints for core tools and MCP tools
+     * - Core tools (web_search, execute_code, etc.) are stable → cache first
+     * - MCP tools (have '_mcp_' in name) are dynamic → cache separately after
+     * - This allows core tools to stay cached when MCP selection changes
+     *
+     * NOTE: Only Claude models support cachePoint - Nova and other models will reject it.
+     */
+    invocationParams(options) {
+        const params = super.invocationParams(options);
+        // Add cachePoint to tools array if promptCache is enabled and tools exist
+        // Only Claude models support cachePoint - check model name
+        const modelId = this.model.toLowerCase();
+        const isClaudeModel = modelId.includes('claude') || modelId.includes('anthropic');
+        if (this.promptCache &&
+            isClaudeModel &&
+            params.toolConfig?.tools &&
+            Array.isArray(params.toolConfig.tools) &&
+            params.toolConfig.tools.length > 0) {
+            // Separate core tools from MCP tools
+            // MCP tools have '_mcp_' in their name (e.g., 'search_emails_mcp_Google-Workspace')
+            const coreTools = [];
+            const mcpTools = [];
+            for (const tool of params.toolConfig.tools) {
+                // Check if tool has a name property with '_mcp_' pattern
+                const toolName = tool.toolSpec?.name ?? '';
+                if (toolName.includes('_mcp_')) {
+                    mcpTools.push(tool);
+                }
+                else {
+                    coreTools.push(tool);
+                }
+            }
+            // Build tools array with strategic cachePoints:
+            // [CoreTool1, CoreTool2, cachePoint] + [MCPTool1, MCPTool2, cachePoint]
+            const toolsWithCache = [];
+            // Add core tools with cachePoint (if any)
+            if (coreTools.length > 0) {
+                toolsWithCache.push(...coreTools);
+                toolsWithCache.push({ cachePoint: { type: 'default' } });
+            }
+            // Add MCP tools with their own cachePoint (if any)
+            if (mcpTools.length > 0) {
+                toolsWithCache.push(...mcpTools);
+                toolsWithCache.push({ cachePoint: { type: 'default' } });
+            }
+            // If no tools at all (shouldn't happen but safety check)
+            if (toolsWithCache.length === 0) {
+                toolsWithCache.push({ cachePoint: { type: 'default' } });
+            }
+            params.toolConfig.tools = toolsWithCache;
+        }
+        // Add serviceTier support
+        const serviceTierType = options?.serviceTier ?? this.serviceTier;
+        return {
+            ...params,
+            serviceTier: serviceTierType ? { type: serviceTierType } : undefined,
+        };
+    }
+    /**
+     * Override _generateNonStreaming to use applicationInferenceProfile as modelId.
+     * Uses the same model-swapping pattern as streaming for consistency.
+     */
+    async _generateNonStreaming(messages, options, runManager) {
+        // Temporarily swap model for applicationInferenceProfile support
+        const originalModel = this.model;
+        if (this.applicationInferenceProfile != null &&
+            this.applicationInferenceProfile !== '') {
+            this.model = this.applicationInferenceProfile;
+        }
+        try {
+            return await super._generateNonStreaming(messages, options, runManager);
+        }
+        finally {
+            // Restore original model
+            this.model = originalModel;
+        }
+    }
+    /**
+     * Override _streamResponseChunks to:
+     * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)
+     * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts
+     * 3. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata
+     *
+     * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's
+     * internal chunk handling which correctly preserves array content for reasoning blocks.
+     */
+    async *_streamResponseChunks(messages, options, runManager) {
+        // Temporarily swap model for applicationInferenceProfile support
+        const originalModel = this.model;
+        if (this.applicationInferenceProfile != null &&
+            this.applicationInferenceProfile !== '') {
+            this.model = this.applicationInferenceProfile;
+        }
+        try {
+            // Use parent's streaming logic which correctly handles reasoning content
+            const baseStream = super._streamResponseChunks(messages, options, runManager);
+            for await (const chunk of baseStream) {
+                // Clean and enhance chunk
+                yield this.processChunk(chunk);
+            }
+        }
+        finally {
+            // Restore original model
+            this.model = originalModel;
+        }
+    }
+    /**
+     * Process a chunk by:
+     * 1. Removing contentBlockIndex from response_metadata
+     * 2. Extracting cache token information from Bedrock's usage data
+     */
+    processChunk(chunk) {
+        const message = chunk.message;
+        if (!(message instanceof messages.AIMessageChunk)) {
+            return chunk;
+        }
+        const responseMetadata = message.response_metadata;
+        let needsModification = false;
+        let cleanedMetadata = responseMetadata;
+        // Check if contentBlockIndex exists anywhere in response_metadata
+        const hasContentBlockIndex = this.hasContentBlockIndex(responseMetadata);
+        if (hasContentBlockIndex) {
+            cleanedMetadata = this.removeContentBlockIndex(responseMetadata);
+            needsModification = true;
+        }
+        // Extract cache tokens from metadata.usage (Bedrock streaming format)
+        // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens
+        const metadata = responseMetadata.metadata;
+        const usage = (metadata?.usage ?? responseMetadata.usage);
+        let enhancedUsageMetadata = message.usage_metadata;
+        if (usage) {
+            const cacheRead = usage.cacheReadInputTokens ?? 0;
+            const cacheWrite = usage.cacheWriteInputTokens ?? 0;
+            const inputTokens = usage.inputTokens ?? 0;
+            const outputTokens = usage.outputTokens ?? 0;
+            if (cacheRead > 0 || cacheWrite > 0) {
+                needsModification = true;
+                enhancedUsageMetadata = {
+                    input_tokens: message.usage_metadata?.input_tokens ?? inputTokens,
+                    output_tokens: message.usage_metadata?.output_tokens ?? outputTokens,
+                    total_tokens: message.usage_metadata?.total_tokens ??
+                        usage.totalTokens ??
+                        0,
+                    input_token_details: {
+                        cache_read: cacheRead,
+                        cache_creation: cacheWrite,
+                    },
+                };
+            }
+        }
+        if (needsModification) {
+            return new outputs.ChatGenerationChunk({
+                text: chunk.text,
+                message: new messages.AIMessageChunk({
+                    ...message,
+                    response_metadata: cleanedMetadata,
+                    usage_metadata: enhancedUsageMetadata,
+                }),
+                generationInfo: chunk.generationInfo,
+            });
+        }
+        return chunk;
+    }
+    /**
+     * Check if contentBlockIndex exists at any level in the object
+     */
+    hasContentBlockIndex(obj) {
+        if (obj === null || obj === undefined || typeof obj !== 'object') {
+            return false;
+        }
+        if ('contentBlockIndex' in obj) {
+            return true;
+        }
+        for (const value of Object.values(obj)) {
+            if (typeof value === 'object' && value !== null) {
+                if (this.hasContentBlockIndex(value)) {
+                    return true;
+                }
+            }
+        }
+        return false;
+    }
+    /**
+     * Recursively remove contentBlockIndex from all levels of an object
+     */
+    removeContentBlockIndex(obj) {
+        if (obj === null || obj === undefined) {
+            return obj;
+        }
+        if (Array.isArray(obj)) {
+            return obj.map((item) => this.removeContentBlockIndex(item));
+        }
+        if (typeof obj === 'object') {
+            const cleaned = {};
+            for (const [key, value] of Object.entries(obj)) {
+                if (key !== 'contentBlockIndex') {
+                    cleaned[key] = this.removeContentBlockIndex(value);
+                }
+            }
+            return cleaned;
+        }
+        return obj;
+    }
+}
+
+exports.CustomChatBedrockConverse = CustomChatBedrockConverse;
+//# sourceMappingURL=index.cjs.map

package/dist/cjs/llm/bedrock/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.cjs","sources":["../../../../src/llm/bedrock/index.ts"],"sourcesContent":["/**\n * Optimized ChatBedrockConverse wrapper that fixes contentBlockIndex conflicts\n * and adds support for:\n *\n * - Prompt caching support for Bedrock Converse API (Illuma feature)\n * - Application Inference Profiles (PR #9129)\n * - Service Tiers (Priority/Standard/Flex) (PR #9785) - requires AWS SDK 3.966.0+\n *\n * Bedrock sends the same contentBlockIndex for both text and tool_use content blocks,\n * causing LangChain's merge logic to fail with \"field[contentBlockIndex] already exists\"\n * errors. This wrapper simply strips contentBlockIndex from response_metadata to avoid\n * the conflict.\n *\n * The contentBlockIndex field is only used internally by Bedrock's streaming protocol\n * and isn't needed by application logic - the index field on tool_call_chunks serves\n * the purpose of tracking tool call ordering.\n *\n * PROMPT CACHING:\n * When promptCache: true is set, this wrapper adds cachePoint markers to the tools array\n * to enable Bedrock prompt caching for tool definitions. This allows tool schemas to be\n * cached and reused across requests, reducing latency and costs.\n *\n * CACHE TOKEN EXTRACTION:\n * LangChain AWS doesn't extract cacheReadInputTokens/cacheWriteInputTokens from Bedrock's\n * response. This wrapper adds input_token_details to usage_metadata with cache information.\n */\n\nimport { ChatBedrockConverse } from '@langchain/aws';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { BaseMessage, UsageMetadata } from '@langchain/core/messages';\nimport { ChatGenerationChunk, ChatResult } from '@langchain/core/outputs';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { ChatBedrockConverseInput } from '@langchain/aws';\n\n/**\n * Service tier type for Bedrock invocations.\n * Requires AWS SDK >= 3.966.0 to actually work.\n * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n */\nexport type ServiceTierType = 'priority' | 'default' | 'flex' | 'reserved';\n\n/**\n * Extended input interface with additional features:\n * - promptCache: Enable Bedrock prompt caching for tool definitions\n * - applicationInferenceProfile: Use an inference profile ARN instead of model ID\n * - serviceTier: Specify service tier (Priority, Standard, Flex, Reserved)\n */\nexport interface CustomChatBedrockConverseInput\n  extends ChatBedrockConverseInput {\n  /**\n   * Enable Bedrock prompt caching for tool definitions.\n   * When true, adds cachePoint markers to tools array.\n   */\n  promptCache?: boolean;\n\n  /**\n   * Application Inference Profile ARN to use for the model.\n   * For example, \"arn:aws:bedrock:eu-west-1:123456789102:application-inference-profile/fm16bt65tzgx\"\n   * When provided, this ARN will be used for the actual inference calls instead of the model ID.\n   * Must still provide `model` as normal modelId to benefit from all the metadata.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles-create.html\n   */\n  applicationInferenceProfile?: string;\n\n  /**\n   * Service tier for model invocation.\n   * Specifies the processing tier type used for serving the request.\n   * Supported values are 'priority', 'default', 'flex', and 'reserved'.\n   *\n   * - 'priority': Prioritized processing for lower latency\n   * - 'default': Standard processing tier\n   * - 'flex': Flexible processing tier with lower cost\n   * - 'reserved': Reserved capacity for consistent performance\n   *\n   * If not provided, AWS uses the default tier.\n   * Note: Requires AWS SDK >= 3.966.0 to work.\n   * @see https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html\n   */\n  serviceTier?: ServiceTierType;\n}\n\n/**\n * Extended call options with serviceTier override support.\n */\nexport interface CustomChatBedrockConverseCallOptions {\n  serviceTier?: ServiceTierType;\n}\n\nexport class CustomChatBedrockConverse extends ChatBedrockConverse {\n  /** Enable Bedrock prompt caching for tool definitions */\n  promptCache: boolean;\n\n  /** Application Inference Profile ARN to use instead of model ID */\n  applicationInferenceProfile?: string;\n\n  /** Service tier for model invocation */\n  serviceTier?: ServiceTierType;\n\n  constructor(fields?: CustomChatBedrockConverseInput) {\n    super(fields);\n    this.promptCache = fields?.promptCache ?? false;\n    this.applicationInferenceProfile = fields?.applicationInferenceProfile;\n    this.serviceTier = fields?.serviceTier;\n\n    // Fix: Force supportsToolChoiceValues for Claude models\n    // The parent constructor checks `model.includes('claude-3')` but this fails when:\n    // 1. Using applicationInferenceProfile ARNs (arn:aws:bedrock:...)\n    // 2. Using different naming conventions (claude-4, claude-opus-4, etc.)\n    // We need to ensure tool_choice is properly set for withStructuredOutput to work\n    const modelName = (fields?.model ?? '').toLowerCase();\n    const profileName = (fields?.applicationInferenceProfile ?? '').toLowerCase();\n    const isClaudeModel =\n      modelName.includes('claude') ||\n      modelName.includes('anthropic') ||\n      profileName.includes('claude') ||\n      profileName.includes('anthropic');\n\n    if (isClaudeModel && !this.supportsToolChoiceValues?.length) {\n      // Claude models support all tool choice values\n      this.supportsToolChoiceValues = ['auto', 'any', 'tool'];\n    }\n  }\n\n  static lc_name(): string {\n    return 'IllumaBedrockConverse';\n  }\n\n  /**\n   * Get the model ID to use for API calls.\n   * Returns applicationInferenceProfile if set, otherwise returns this.model.\n   */\n  protected getModelId(): string {\n    return this.applicationInferenceProfile ?? this.model;\n  }\n\n  /**\n   * Override invocationParams to:\n   * 1. Add cachePoint to tools when promptCache is enabled\n   * 2. Add serviceTier support\n   *\n   * CACHING STRATEGY: Separate cachePoints for core tools and MCP tools\n   * - Core tools (web_search, execute_code, etc.) are stable → cache first\n   * - MCP tools (have '_mcp_' in name) are dynamic → cache separately after\n   * - This allows core tools to stay cached when MCP selection changes\n   *\n   * NOTE: Only Claude models support cachePoint - Nova and other models will reject it.\n   */\n  override invocationParams(\n    options?: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions\n  ): ReturnType<ChatBedrockConverse['invocationParams']> & {\n    serviceTier?: { type: ServiceTierType };\n  } {\n    const params = super.invocationParams(options);\n\n    // Add cachePoint to tools array if promptCache is enabled and tools exist\n    // Only Claude models support cachePoint - check model name\n    const modelId = this.model.toLowerCase();\n    const isClaudeModel =\n      modelId.includes('claude') || modelId.includes('anthropic');\n\n    if (\n      this.promptCache &&\n      isClaudeModel &&\n      params.toolConfig?.tools &&\n      Array.isArray(params.toolConfig.tools) &&\n      params.toolConfig.tools.length > 0\n    ) {\n      // Separate core tools from MCP tools\n      // MCP tools have '_mcp_' in their name (e.g., 'search_emails_mcp_Google-Workspace')\n      const coreTools: typeof params.toolConfig.tools = [];\n      const mcpTools: typeof params.toolConfig.tools = [];\n\n      for (const tool of params.toolConfig.tools) {\n        // Check if tool has a name property with '_mcp_' pattern\n        const toolName =\n          (tool as { toolSpec?: { name?: string } }).toolSpec?.name ?? '';\n        if (toolName.includes('_mcp_')) {\n          mcpTools.push(tool);\n        } else {\n          coreTools.push(tool);\n        }\n      }\n\n      // Build tools array with strategic cachePoints:\n      // [CoreTool1, CoreTool2, cachePoint] + [MCPTool1, MCPTool2, cachePoint]\n      const toolsWithCache: typeof params.toolConfig.tools = [];\n\n      // Add core tools with cachePoint (if any)\n      if (coreTools.length > 0) {\n        toolsWithCache.push(...coreTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // Add MCP tools with their own cachePoint (if any)\n      if (mcpTools.length > 0) {\n        toolsWithCache.push(...mcpTools);\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      // If no tools at all (shouldn't happen but safety check)\n      if (toolsWithCache.length === 0) {\n        toolsWithCache.push({ cachePoint: { type: 'default' } });\n      }\n\n      params.toolConfig.tools = toolsWithCache;\n    }\n\n    // Add serviceTier support\n    const serviceTierType = options?.serviceTier ?? this.serviceTier;\n\n    return {\n      ...params,\n      serviceTier: serviceTierType ? { type: serviceTierType } : undefined,\n    };\n  }\n\n  /**\n   * Override _generateNonStreaming to use applicationInferenceProfile as modelId.\n   * Uses the same model-swapping pattern as streaming for consistency.\n   */\n  override async _generateNonStreaming(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): Promise<ChatResult> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      return await super._generateNonStreaming(messages, options, runManager);\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Override _streamResponseChunks to:\n   * 1. Use applicationInferenceProfile as modelId (by temporarily swapping this.model)\n   * 2. Strip contentBlockIndex from response_metadata to prevent merge conflicts\n   * 3. Extract cacheReadInputTokens/cacheWriteInputTokens and add to usage_metadata\n   *\n   * Note: We delegate to super._streamResponseChunks() to preserve @langchain/aws's\n   * internal chunk handling which correctly preserves array content for reasoning blocks.\n   */\n  override async *_streamResponseChunks(\n    messages: BaseMessage[],\n    options: this['ParsedCallOptions'] & CustomChatBedrockConverseCallOptions,\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    // Temporarily swap model for applicationInferenceProfile support\n    const originalModel = this.model;\n    if (\n      this.applicationInferenceProfile != null &&\n      this.applicationInferenceProfile !== ''\n    ) {\n      this.model = this.applicationInferenceProfile;\n    }\n\n    try {\n      // Use parent's streaming logic which correctly handles reasoning content\n      const baseStream = super._streamResponseChunks(\n        messages,\n        options,\n        runManager\n      );\n\n      for await (const chunk of baseStream) {\n        // Clean and enhance chunk\n        yield this.processChunk(chunk);\n      }\n    } finally {\n      // Restore original model\n      this.model = originalModel;\n    }\n  }\n\n  /**\n   * Process a chunk by:\n   * 1. Removing contentBlockIndex from response_metadata\n   * 2. Extracting cache token information from Bedrock's usage data\n   */\n  private processChunk(chunk: ChatGenerationChunk): ChatGenerationChunk {\n    const message = chunk.message;\n    if (!(message instanceof AIMessageChunk)) {\n      return chunk;\n    }\n\n    const responseMetadata = message.response_metadata as Record<\n      string,\n      unknown\n    >;\n    let needsModification = false;\n    let cleanedMetadata = responseMetadata;\n\n    // Check if contentBlockIndex exists anywhere in response_metadata\n    const hasContentBlockIndex = this.hasContentBlockIndex(responseMetadata);\n    if (hasContentBlockIndex) {\n      cleanedMetadata = this.removeContentBlockIndex(\n        responseMetadata\n      ) as Record<string, unknown>;\n      needsModification = true;\n    }\n\n    // Extract cache tokens from metadata.usage (Bedrock streaming format)\n    // The metadata chunk contains usage with cacheReadInputTokens/cacheWriteInputTokens\n    const metadata = responseMetadata.metadata as\n      | Record<string, unknown>\n      | undefined;\n    const usage = (metadata?.usage ?? responseMetadata.usage) as\n      | Record<string, unknown>\n      | undefined;\n\n    let enhancedUsageMetadata: UsageMetadata | undefined =\n      message.usage_metadata;\n\n    if (usage) {\n      const cacheRead = (usage.cacheReadInputTokens as number | undefined) ?? 0;\n      const cacheWrite =\n        (usage.cacheWriteInputTokens as number | undefined) ?? 0;\n      const inputTokens = (usage.inputTokens as number | undefined) ?? 0;\n      const outputTokens = (usage.outputTokens as number | undefined) ?? 0;\n\n      if (cacheRead > 0 || cacheWrite > 0) {\n        needsModification = true;\n        enhancedUsageMetadata = {\n          input_tokens: message.usage_metadata?.input_tokens ?? inputTokens,\n          output_tokens: message.usage_metadata?.output_tokens ?? outputTokens,\n          total_tokens:\n            message.usage_metadata?.total_tokens ??\n            (usage.totalTokens as number | undefined) ??\n            0,\n          input_token_details: {\n            cache_read: cacheRead,\n            cache_creation: cacheWrite,\n          },\n        };\n      }\n    }\n\n    if (needsModification) {\n      return new ChatGenerationChunk({\n        text: chunk.text,\n        message: new AIMessageChunk({\n          ...message,\n          response_metadata: cleanedMetadata,\n          usage_metadata: enhancedUsageMetadata,\n        }),\n        generationInfo: chunk.generationInfo,\n      });\n    }\n\n    return chunk;\n  }\n\n  /**\n   * Check if contentBlockIndex exists at any level in the object\n   */\n  private hasContentBlockIndex(obj: unknown): boolean {\n    if (obj === null || obj === undefined || typeof obj !== 'object') {\n      return false;\n    }\n\n    if ('contentBlockIndex' in obj) {\n      return true;\n    }\n\n    for (const value of Object.values(obj)) {\n      if (typeof value === 'object' && value !== null) {\n        if (this.hasContentBlockIndex(value)) {\n          return true;\n        }\n      }\n    }\n\n    return false;\n  }\n\n  /**\n   * Recursively remove contentBlockIndex from all levels of an object\n   */\n  private removeContentBlockIndex(obj: unknown): unknown {\n    if (obj === null || obj === undefined) {\n      return obj;\n    }\n\n    if (Array.isArray(obj)) {\n      return obj.map((item) => this.removeContentBlockIndex(item));\n    }\n\n    if (typeof obj === 'object') {\n      const cleaned: Record<string, unknown> = {};\n      for (const [key, value] of Object.entries(obj)) {\n        if (key !== 'contentBlockIndex') {\n          cleaned[key] = this.removeContentBlockIndex(value);\n        }\n      }\n      return cleaned;\n    }\n\n    return obj;\n  }\n}\n\nexport type { ChatBedrockConverseInput };\n"],"names":["ChatBedrockConverse","AIMessageChunk","ChatGenerationChunk"],"mappings":";;;;;;AAAA;;;;;;;;;;;;;;;;;;;;;;;;;AAyBG;AA+DG,MAAO,yBAA0B,SAAQA,uBAAmB,CAAA;;AAEhE,IAAA,WAAW;;AAGX,IAAA,2BAA2B;;AAG3B,IAAA,WAAW;AAEX,IAAA,WAAA,CAAY,MAAuC,EAAA;QACjD,KAAK,CAAC,MAAM,CAAC;QACb,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW,IAAI,KAAK;AAC/C,QAAA,IAAI,CAAC,2BAA2B,GAAG,MAAM,EAAE,2BAA2B;AACtE,QAAA,IAAI,CAAC,WAAW,GAAG,MAAM,EAAE,WAAW;;;;;;AAOtC,QAAA,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,KAAK,IAAI,EAAE,EAAE,WAAW,EAAE;AACrD,QAAA,MAAM,WAAW,GAAG,CAAC,MAAM,EAAE,2BAA2B,IAAI,EAAE,EAAE,WAAW,EAAE;AAC7E,QAAA,MAAM,aAAa,GACjB,SAAS,CAAC,QAAQ,CAAC,QAAQ,CAAC;AAC5B,YAAA,SAAS,CAAC,QAAQ,CAAC,WAAW,CAAC;AAC/B,YAAA,WAAW,CAAC,QAAQ,CAAC,QAAQ,CAAC;AAC9B,YAAA,WAAW,CAAC,QAAQ,CAAC,WAAW,CAAC;QAEnC,IAAI,aAAa,IAAI,CAAC,IAAI,CAAC,wBAAwB,EAAE,MAAM,EAAE;;YAE3D,IAAI,CAAC,wBAAwB,GAAG,CAAC,MAAM,EAAE,KAAK,EAAE,MAAM,CAAC;;;AAI3D,IAAA,OAAO,OAAO,GAAA;AACZ,QAAA,OAAO,uBAAuB;;AAGhC;;;AAGG;IACO,UAAU,GAAA;AAClB,QAAA,OAAO,IAAI,CAAC,2BAA2B,IAAI,IAAI,CAAC,KAAK;;AAGvD;;;;;;;;;;;AAWG;AACM,IAAA,gBAAgB,CACvB,OAA0E,EAAA;QAI1E,MAAM,MAAM,GAAG,KAAK,CAAC,gBAAgB,CAAC,OAAO,CAAC;;;QAI9C,MAAM,OAAO,GAAG,IAAI,CAAC,KAAK,CAAC,WAAW,EAAE;AACxC,QAAA,MAAM,aAAa,GACjB,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAC;QAE7D,IACE,IAAI,CAAC,WAAW;YAChB,aAAa;YACb,MAAM,CAAC,UAAU,EAAE,KAAK;YACxB,KAAK,CAAC,OAAO,CAAC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC;YACtC,MAAM,CAAC,UAAU,CAAC,KAAK,CAAC,MAAM,GAAG,CAAC,EAClC;;;YAGA,MAAM,SAAS,GAAmC,EAAE;YACpD,MAAM,QAAQ,GAAmC,EAAE;YAEnD,KAAK,MAAM,IAAI,IAAI,MAAM,CAAC,UAAU,CAAC,KAAK,EAAE;;gBAE1C,MAAM,QAAQ,GACX,IAAyC,CAAC,QAAQ,EAAE,IAAI,IAAI,EAAE;AACjE,gBAAA,IAAI,QAAQ,CAAC,QAAQ,CAAC,OAAO,CAAC,EAAE;AAC9B,oBAAA,QAAQ,CAAC,IAAI,CAAC,IAAI,CAAC;;qBACd;AACL,oBAAA,SAAS,CAAC,IAAI,CAAC,IAAI,CAAC;;;;;YAMxB,MAAM,cAAc,GAAmC,EAAE;;AAGzD,YAAA,IAAI,SAAS,CAAC,MAAM,GAAG,CAAC,EAAE;AACxB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,SAAS,CAAC;AACjC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,QAAQ,CAAC,MAAM,GAAG,CAAC,EAAE;AACvB,gBAAA,cAAc,CAAC,IAAI,CAAC,GAAG,QAAQ,CAAC;AAChC,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;;AAI1D,YAAA,IAAI,cAAc,CAAC,MAAM,KAAK,CAAC,EAAE;AAC/B,gBAAA,cAAc,CAAC,IAAI,CAAC,EAAE,UAAU,EAAE,EAAE,IAAI,EAAE,SAAS,EAAE,EAAE,CAAC;;AAG1D,YAAA,MAAM,CAAC,UAAU,CAAC,KAAK,GAAG,cAAc;;;QAI1C,MAAM,eAAe,GAAG,OAAO,EAAE,WAAW,IAAI,IAAI,CAAC,WAAW;QAEhE,OAAO;AACL,YAAA,GAAG,MAAM;AACT,YAAA,WAAW,EAAE,eAAe,GAAG,EAAE,IAAI,EAAE,eAAe,EAAE,GAAG,SAAS;SACrE;;AAGH;;;AAGG;AACM,IAAA,MAAM,qBAAqB,CAClC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;YACF,OAAO,MAAM,KAAK,CAAC,qBAAqB,CAAC,QAAQ,EAAE,OAAO,EAAE,UAAU,CAAC;;gBAC/D;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;;;;;;;AAQG;IACM,OAAO,qBAAqB,CACnC,QAAuB,EACvB,OAAyE,EACzE,UAAqC,EAAA;;AAGrC,QAAA,MAAM,aAAa,GAAG,IAAI,CAAC,KAAK;AAChC,QAAA,IACE,IAAI,CAAC,2BAA2B,IAAI,IAAI;AACxC,YAAA,IAAI,CAAC,2BAA2B,KAAK,EAAE,EACvC;AACA,YAAA,IAAI,CAAC,KAAK,GAAG,IAAI,CAAC,2BAA2B;;AAG/C,QAAA,IAAI;;AAEF,YAAA,MAAM,UAAU,GAAG,KAAK,CAAC,qBAAqB,CAC5C,QAAQ,EACR,OAAO,EACP,UAAU,CACX;AAED,YAAA,WAAW,MAAM,KAAK,IAAI,UAAU,EAAE;;AAEpC,gBAAA,MAAM,IAAI,CAAC,YAAY,CAAC,KAAK,CAAC;;;gBAExB;;AAER,YAAA,IAAI,CAAC,KAAK,GAAG,aAAa;;;AAI9B;;;;AAIG;AACK,IAAA,YAAY,CAAC,KAA0B,EAAA;AAC7C,QAAA,MAAM,OAAO,GAAG,KAAK,CAAC,OAAO;AAC7B,QAAA,IAAI,EAAE,OAAO,YAAYC,uBAAc,CAAC,EAAE;AACxC,YAAA,OAAO,KAAK;;AAGd,QAAA,MAAM,gBAAgB,GAAG,OAAO,CAAC,iBAGhC;QACD,IAAI,iBAAiB,GAAG,KAAK;QAC7B,IAAI,eAAe,GAAG,gBAAgB;;QAGtC,MAAM,oBAAoB,GAAG,IAAI,CAAC,oBAAoB,CAAC,gBAAgB,CAAC;QACxE,IAAI,oBAAoB,EAAE;AACxB,YAAA,eAAe,GAAG,IAAI,CAAC,uBAAuB,CAC5C,gBAAgB,CACU;YAC5B,iBAAiB,GAAG,IAAI;;;;AAK1B,QAAA,MAAM,QAAQ,GAAG,gBAAgB,CAAC,QAErB;QACb,MAAM,KAAK,IAAI,QAAQ,EAAE,KAAK,IAAI,gBAAgB,CAAC,KAAK,CAE3C;AAEb,QAAA,IAAI,qBAAqB,GACvB,OAAO,CAAC,cAAc;QAExB,IAAI,KAAK,EAAE;AACT,YAAA,MAAM,SAAS,GAAI,KAAK,CAAC,oBAA2C,IAAI,CAAC;AACzE,YAAA,MAAM,UAAU,GACb,KAAK,CAAC,qBAA4C,IAAI,CAAC;AAC1D,YAAA,MAAM,WAAW,GAAI,KAAK,CAAC,WAAkC,IAAI,CAAC;AAClE,YAAA,MAAM,YAAY,GAAI,KAAK,CAAC,YAAmC,IAAI,CAAC;YAEpE,IAAI,SAAS,GAAG,CAAC,IAAI,UAAU,GAAG,CAAC,EAAE;gBACnC,iBAAiB,GAAG,IAAI;AACxB,gBAAA,qBAAqB,GAAG;AACtB,oBAAA,YAAY,EAAE,OAAO,CAAC,cAAc,EAAE,YAAY,IAAI,WAAW;AACjE,oBAAA,aAAa,EAAE,OAAO,CAAC,cAAc,EAAE,aAAa,IAAI,YAAY;AACpE,oBAAA,YAAY,EACV,OAAO,CAAC,cAAc,EAAE,YAAY;AACnC,wBAAA,KAAK,CAAC,WAAkC;wBACzC,CAAC;AACH,oBAAA,mBAAmB,EAAE;AACnB,wBAAA,UAAU,EAAE,SAAS;AACrB,wBAAA,cAAc,EAAE,UAAU;AAC3B,qBAAA;iBACF;;;QAIL,IAAI,iBAAiB,EAAE;YACrB,OAAO,IAAIC,2BAAmB,CAAC;gBAC7B,IAAI,EAAE,KAAK,CAAC,IAAI;gBAChB,OAAO,EAAE,IAAID,uBAAc,CAAC;AAC1B,oBAAA,GAAG,OAAO;AACV,oBAAA,iBAAiB,EAAE,eAAe;AAClC,oBAAA,cAAc,EAAE,qBAAqB;iBACtC,CAAC;gBACF,cAAc,EAAE,KAAK,CAAC,cAAc;AACrC,aAAA,CAAC;;AAGJ,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,oBAAoB,CAAC,GAAY,EAAA;AACvC,QAAA,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;AAChE,YAAA,OAAO,KAAK;;AAGd,QAAA,IAAI,mBAAmB,IAAI,GAAG,EAAE;AAC9B,YAAA,OAAO,IAAI;;QAGb,KAAK,MAAM,KAAK,IAAI,MAAM,CAAC,MAAM,CAAC,GAAG,CAAC,EAAE;YACtC,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,EAAE;AAC/C,gBAAA,IAAI,IAAI,CAAC,oBAAoB,CAAC,KAAK,CAAC,EAAE;AACpC,oBAAA,OAAO,IAAI;;;;AAKjB,QAAA,OAAO,KAAK;;AAGd;;AAEG;AACK,IAAA,uBAAuB,CAAC,GAAY,EAAA;QAC1C,IAAI,GAAG,KAAK,IAAI,IAAI,GAAG,KAAK,SAAS,EAAE;AACrC,YAAA,OAAO,GAAG;;AAGZ,QAAA,IAAI,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AACtB,YAAA,OAAO,GAAG,CAAC,GAAG,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,uBAAuB,CAAC,IAAI,CAAC,CAAC;;AAG9D,QAAA,IAAI,OAAO,GAAG,KAAK,QAAQ,EAAE;YAC3B,MAAM,OAAO,GAA4B,EAAE;AAC3C,YAAA,KAAK,MAAM,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,GAAG,CAAC,EAAE;AAC9C,gBAAA,IAAI,GAAG,KAAK,mBAAmB,EAAE;oBAC/B,OAAO,CAAC,GAAG,CAAC,GAAG,IAAI,CAAC,uBAAuB,CAAC,KAAK,CAAC;;;AAGtD,YAAA,OAAO,OAAO;;AAGhB,QAAA,OAAO,GAAG;;AAEb;;;;"}

package/dist/cjs/llm/fake.cjs

@@ -0,0 +1,97 @@
+'use strict';
+
+var outputs = require('@langchain/core/outputs');
+var messages = require('@langchain/core/messages');
+var testing = require('@langchain/core/utils/testing');
+
+class FakeChatModel extends testing.FakeListChatModel {
+    splitStrategy;
+    toolCalls = [];
+    addedToolCalls = false;
+    constructor({ responses, sleep, emitCustomEvent, splitStrategy = { type: 'regex', value: /(?<=\s+)|(?=\s+)/ }, toolCalls = [] }) {
+        super({ responses, sleep, emitCustomEvent });
+        this.splitStrategy = splitStrategy;
+        this.toolCalls = toolCalls;
+    }
+    splitText(text) {
+        if (this.splitStrategy.type === 'regex') {
+            return text.split(this.splitStrategy.value);
+        }
+        else {
+            const chunkSize = this.splitStrategy.value;
+            const chunks = [];
+            for (let i = 0; i < text.length; i += chunkSize) {
+                chunks.push(text.slice(i, i + chunkSize));
+            }
+            return chunks;
+        }
+    }
+    _createResponseChunk(text, tool_call_chunks) {
+        return new outputs.ChatGenerationChunk({
+            text,
+            generationInfo: {},
+            message: new messages.AIMessageChunk({
+                content: text,
+                tool_call_chunks,
+                additional_kwargs: tool_call_chunks ? {
+                    tool_calls: tool_call_chunks.map((toolCall) => ({
+                        index: toolCall.index ?? 0,
+                        id: toolCall.id ?? '',
+                        type: 'function',
+                        function: {
+                            name: toolCall.name ?? '',
+                            arguments: toolCall.args ?? '',
+                        },
+                    })),
+                } : undefined,
+            })
+        });
+    }
+    async *_streamResponseChunks(_messages, options, runManager) {
+        const response = this._currentResponse();
+        this._incrementResponse();
+        if (this.emitCustomEvent) {
+            await runManager?.handleCustomEvent('some_test_event', {
+                someval: true,
+            });
+        }
+        const chunks = this.splitText(response);
+        for await (const chunk of chunks) {
+            await this._sleepIfRequested();
+            if (options.thrownErrorString != null && options.thrownErrorString) {
+                throw new Error(options.thrownErrorString);
+            }
+            const responseChunk = super._createResponseChunk(chunk);
+            yield responseChunk;
+            void runManager?.handleLLMNewToken(chunk);
+        }
+        await this._sleepIfRequested();
+        if (this.toolCalls.length > 0 && !this.addedToolCalls) {
+            this.addedToolCalls = true;
+            const toolCallChunks = this.toolCalls.map((toolCall) => {
+                return {
+                    name: toolCall.name,
+                    args: JSON.stringify(toolCall.args),
+                    id: toolCall.id,
+                    type: 'tool_call_chunk',
+                };
+            });
+            const responseChunk = this._createResponseChunk('', toolCallChunks);
+            yield responseChunk;
+            void runManager?.handleLLMNewToken('');
+        }
+    }
+}
+function createFakeStreamingLLM({ responses, sleep, splitStrategy, toolCalls, }) {
+    return new FakeChatModel({
+        sleep,
+        responses,
+        emitCustomEvent: true,
+        splitStrategy,
+        toolCalls,
+    });
+}
+
+exports.FakeChatModel = FakeChatModel;
+exports.createFakeStreamingLLM = createFakeStreamingLLM;
+//# sourceMappingURL=fake.cjs.map

package/dist/cjs/llm/fake.cjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"fake.cjs","sources":["../../../src/llm/fake.ts"],"sourcesContent":["import { ChatGenerationChunk } from '@langchain/core/outputs';\nimport { AIMessageChunk } from '@langchain/core/messages';\nimport type { CallbackManagerForLLMRun } from '@langchain/core/callbacks/manager';\nimport type { BaseMessage } from '@langchain/core/messages';\nimport { FakeListChatModel } from '@langchain/core/utils/testing';\nimport { ToolCall, ToolCallChunk } from '@langchain/core/messages/tool';\n\ntype SplitStrategy = {\n  type: 'regex' | 'fixed';\n  value: RegExp | number;\n};\n\nexport class FakeChatModel extends FakeListChatModel {\n  private splitStrategy: SplitStrategy;\n  private toolCalls: ToolCall[] = [];\n  private addedToolCalls: boolean = false;\n\n  constructor({\n    responses,\n    sleep,\n    emitCustomEvent,\n    splitStrategy = { type: 'regex', value: /(?<=\\s+)|(?=\\s+)/ },\n    toolCalls = []\n  }: {\n    responses: string[];\n    sleep?: number;\n    emitCustomEvent?: boolean;\n    splitStrategy?: SplitStrategy;\n    toolCalls?: ToolCall[];\n  }) {\n    super({ responses, sleep, emitCustomEvent });\n    this.splitStrategy = splitStrategy;\n    this.toolCalls = toolCalls;\n  }\n\n  private splitText(text: string): string[] {\n    if (this.splitStrategy.type === 'regex') {\n      return text.split(this.splitStrategy.value as RegExp);\n    } else {\n      const chunkSize = this.splitStrategy.value as number;\n      const chunks: string[] = [];\n      for (let i = 0; i < text.length; i += chunkSize) {\n        chunks.push(text.slice(i, i + chunkSize));\n      }\n      return chunks;\n    }\n  }\n  _createResponseChunk(text: string, tool_call_chunks?: ToolCallChunk[]): ChatGenerationChunk {\n    return new ChatGenerationChunk({\n      text,\n      generationInfo: {},\n      message: new AIMessageChunk({\n        content: text,\n        tool_call_chunks,\n        additional_kwargs: tool_call_chunks ? {\n          tool_calls: tool_call_chunks.map((toolCall) => ({\n            index: toolCall.index ?? 0,\n            id: toolCall.id ?? '',\n            type: 'function',\n            function: {\n              name: toolCall.name ?? '',\n              arguments: toolCall.args ?? '',\n            },\n          })),\n        } : undefined,\n      })});\n  }\n\n  async *_streamResponseChunks(\n    _messages: BaseMessage[],\n    options: this['ParsedCallOptions'],\n    runManager?: CallbackManagerForLLMRun\n  ): AsyncGenerator<ChatGenerationChunk> {\n    const response = this._currentResponse();\n    this._incrementResponse();\n\n    if (this.emitCustomEvent) {\n      await runManager?.handleCustomEvent('some_test_event', {\n        someval: true,\n      });\n    }\n\n    const chunks = this.splitText(response);\n    for await (const chunk of chunks) {\n      await this._sleepIfRequested();\n\n      if (options.thrownErrorString != null && options.thrownErrorString) {\n        throw new Error(options.thrownErrorString);\n      }\n\n      const responseChunk = super._createResponseChunk(chunk);\n      yield responseChunk;\n      void runManager?.handleLLMNewToken(chunk);\n    }\n\n    await this._sleepIfRequested();\n    if (this.toolCalls.length > 0 && !this.addedToolCalls) {\n      this.addedToolCalls = true;\n      const toolCallChunks = this.toolCalls.map((toolCall) => {;\n        return {\n          name: toolCall.name,\n          args: JSON.stringify(toolCall.args),\n          id: toolCall.id,\n          type: 'tool_call_chunk',\n        } as ToolCallChunk;\n      });\n      const responseChunk = this._createResponseChunk('', toolCallChunks);\n      yield responseChunk;\n      void runManager?.handleLLMNewToken('');\n    }\n  }\n}\n\nexport function createFakeStreamingLLM({\n  responses,\n  sleep,\n  splitStrategy,\n  toolCalls,\n} : {\n  responses: string[],\n  sleep?: number,\n  splitStrategy?: SplitStrategy,\n  toolCalls?: ToolCall[]\n}\n): FakeChatModel {\n  return new FakeChatModel({\n    sleep,\n    responses,\n    emitCustomEvent: true,\n    splitStrategy,\n    toolCalls,\n  });\n}\n"],"names":["FakeListChatModel","ChatGenerationChunk","AIMessageChunk"],"mappings":";;;;;;AAYM,MAAO,aAAc,SAAQA,yBAAiB,CAAA;AAC1C,IAAA,aAAa;IACb,SAAS,GAAe,EAAE;IAC1B,cAAc,GAAY,KAAK;IAEvC,WAAY,CAAA,EACV,SAAS,EACT,KAAK,EACL,eAAe,EACf,aAAa,GAAG,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,kBAAkB,EAAE,EAC5D,SAAS,GAAG,EAAE,EAOf,EAAA;QACC,KAAK,CAAC,EAAE,SAAS,EAAE,KAAK,EAAE,eAAe,EAAE,CAAC;AAC5C,QAAA,IAAI,CAAC,aAAa,GAAG,aAAa;AAClC,QAAA,IAAI,CAAC,SAAS,GAAG,SAAS;;AAGpB,IAAA,SAAS,CAAC,IAAY,EAAA;QAC5B,IAAI,IAAI,CAAC,aAAa,CAAC,IAAI,KAAK,OAAO,EAAE;YACvC,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,aAAa,CAAC,KAAe,CAAC;;aAChD;AACL,YAAA,MAAM,SAAS,GAAG,IAAI,CAAC,aAAa,CAAC,KAAe;YACpD,MAAM,MAAM,GAAa,EAAE;AAC3B,YAAA,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,CAAC,IAAI,SAAS,EAAE;AAC/C,gBAAA,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,GAAG,SAAS,CAAC,CAAC;;AAE3C,YAAA,OAAO,MAAM;;;IAGjB,oBAAoB,CAAC,IAAY,EAAE,gBAAkC,EAAA;QACnE,OAAO,IAAIC,2BAAmB,CAAC;YAC7B,IAAI;AACJ,YAAA,cAAc,EAAE,EAAE;YAClB,OAAO,EAAE,IAAIC,uBAAc,CAAC;AAC1B,gBAAA,OAAO,EAAE,IAAI;gBACb,gBAAgB;AAChB,gBAAA,iBAAiB,EAAE,gBAAgB,GAAG;oBACpC,UAAU,EAAE,gBAAgB,CAAC,GAAG,CAAC,CAAC,QAAQ,MAAM;AAC9C,wBAAA,KAAK,EAAE,QAAQ,CAAC,KAAK,IAAI,CAAC;AAC1B,wBAAA,EAAE,EAAE,QAAQ,CAAC,EAAE,IAAI,EAAE;AACrB,wBAAA,IAAI,EAAE,UAAU;AAChB,wBAAA,QAAQ,EAAE;AACR,4BAAA,IAAI,EAAE,QAAQ,CAAC,IAAI,IAAI,EAAE;AACzB,4BAAA,SAAS,EAAE,QAAQ,CAAC,IAAI,IAAI,EAAE;AAC/B,yBAAA;AACF,qBAAA,CAAC,CAAC;iBACJ,GAAG,SAAS;aACd;AAAE,SAAA,CAAC;;IAGR,OAAO,qBAAqB,CAC1B,SAAwB,EACxB,OAAkC,EAClC,UAAqC,EAAA;AAErC,QAAA,MAAM,QAAQ,GAAG,IAAI,CAAC,gBAAgB,EAAE;QACxC,IAAI,CAAC,kBAAkB,EAAE;AAEzB,QAAA,IAAI,IAAI,CAAC,eAAe,EAAE;AACxB,YAAA,MAAM,UAAU,EAAE,iBAAiB,CAAC,iBAAiB,EAAE;AACrD,gBAAA,OAAO,EAAE,IAAI;AACd,aAAA,CAAC;;QAGJ,MAAM,MAAM,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC;AACvC,QAAA,WAAW,MAAM,KAAK,IAAI,MAAM,EAAE;AAChC,YAAA,MAAM,IAAI,CAAC,iBAAiB,EAAE;YAE9B,IAAI,OAAO,CAAC,iBAAiB,IAAI,IAAI,IAAI,OAAO,CAAC,iBAAiB,EAAE;AAClE,gBAAA,MAAM,IAAI,KAAK,CAAC,OAAO,CAAC,iBAAiB,CAAC;;YAG5C,MAAM,aAAa,GAAG,KAAK,CAAC,oBAAoB,CAAC,KAAK,CAAC;AACvD,YAAA,MAAM,aAAa;AACnB,YAAA,KAAK,UAAU,EAAE,iBAAiB,CAAC,KAAK,CAAC;;AAG3C,QAAA,MAAM,IAAI,CAAC,iBAAiB,EAAE;AAC9B,QAAA,IAAI,IAAI,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,IAAI,CAAC,IAAI,CAAC,cAAc,EAAE;AACrD,YAAA,IAAI,CAAC,cAAc,GAAG,IAAI;YAC1B,MAAM,cAAc,GAAG,IAAI,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC,QAAQ,KAAI;gBACrD,OAAO;oBACL,IAAI,EAAE,QAAQ,CAAC,IAAI;oBACnB,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,QAAQ,CAAC,IAAI,CAAC;oBACnC,EAAE,EAAE,QAAQ,CAAC,EAAE;AACf,oBAAA,IAAI,EAAE,iBAAiB;iBACP;AACpB,aAAC,CAAC;YACF,MAAM,aAAa,GAAG,IAAI,CAAC,oBAAoB,CAAC,EAAE,EAAE,cAAc,CAAC;AACnE,YAAA,MAAM,aAAa;AACnB,YAAA,KAAK,UAAU,EAAE,iBAAiB,CAAC,EAAE,CAAC;;;AAG3C;AAEK,SAAU,sBAAsB,CAAC,EACrC,SAAS,EACT,KAAK,EACL,aAAa,EACb,SAAS,GAMV,EAAA;IAEC,OAAO,IAAI,aAAa,CAAC;QACvB,KAAK;QACL,SAAS;AACT,QAAA,eAAe,EAAE,IAAI;QACrB,aAAa;QACb,SAAS;AACV,KAAA,CAAC;AACJ;;;;;"}

package/dist/cjs/llm/google/index.cjs

@@ -0,0 +1,216 @@
+'use strict';
+
+var messages = require('@langchain/core/messages');
+var outputs = require('@langchain/core/outputs');
+var googleGenai = require('@langchain/google-genai');
+var env = require('@langchain/core/utils/env');
+var generativeAi = require('@google/generative-ai');
+var common = require('./utils/common.cjs');
+
+/* eslint-disable @typescript-eslint/ban-ts-comment */
+class CustomChatGoogleGenerativeAI extends googleGenai.ChatGoogleGenerativeAI {
+    thinkingConfig;
+    /**
+     * Override to add gemini-3 model support for multimodal and function calling thought signatures
+     */
+    get _isMultimodalModel() {
+        return (this.model.startsWith('gemini-1.5') ||
+            this.model.startsWith('gemini-2') ||
+            (this.model.startsWith('gemma-3-') &&
+                !this.model.startsWith('gemma-3-1b')) ||
+            this.model.startsWith('gemini-3'));
+    }
+    constructor(fields) {
+        super(fields);
+        this.model = fields.model.replace(/^models\//, '');
+        this.maxOutputTokens = fields.maxOutputTokens ?? this.maxOutputTokens;
+        if (this.maxOutputTokens != null && this.maxOutputTokens < 0) {
+            throw new Error('`maxOutputTokens` must be a positive integer');
+        }
+        this.temperature = fields.temperature ?? this.temperature;
+        if (this.temperature != null &&
+            (this.temperature < 0 || this.temperature > 2)) {
+            throw new Error('`temperature` must be in the range of [0.0,2.0]');
+        }
+        this.topP = fields.topP ?? this.topP;
+        if (this.topP != null && this.topP < 0) {
+            throw new Error('`topP` must be a positive integer');
+        }
+        if (this.topP != null && this.topP > 1) {
+            throw new Error('`topP` must be below 1.');
+        }
+        this.topK = fields.topK ?? this.topK;
+        if (this.topK != null && this.topK < 0) {
+            throw new Error('`topK` must be a positive integer');
+        }
+        this.stopSequences = fields.stopSequences ?? this.stopSequences;
+        this.apiKey = fields.apiKey ?? env.getEnvironmentVariable('GOOGLE_API_KEY');
+        if (this.apiKey == null || this.apiKey === '') {
+            throw new Error('Please set an API key for Google GenerativeAI ' +
+                'in the environment variable GOOGLE_API_KEY ' +
+                'or in the `apiKey` field of the ' +
+                'ChatGoogleGenerativeAI constructor');
+        }
+        this.safetySettings = fields.safetySettings ?? this.safetySettings;
+        if (this.safetySettings && this.safetySettings.length > 0) {
+            const safetySettingsSet = new Set(this.safetySettings.map((s) => s.category));
+            if (safetySettingsSet.size !== this.safetySettings.length) {
+                throw new Error('The categories in `safetySettings` array must be unique');
+            }
+        }
+        this.thinkingConfig = fields.thinkingConfig ?? this.thinkingConfig;
+        this.streaming = fields.streaming ?? this.streaming;
+        this.json = fields.json;
+        // @ts-ignore - Accessing private property from parent class
+        this.client = new generativeAi.GoogleGenerativeAI(this.apiKey).getGenerativeModel({
+            model: this.model,
+            safetySettings: this.safetySettings,
+            generationConfig: {
+                stopSequences: this.stopSequences,
+                maxOutputTokens: this.maxOutputTokens,
+                temperature: this.temperature,
+                topP: this.topP,
+                topK: this.topK,
+                ...(this.json != null
+                    ? { responseMimeType: 'application/json' }
+                    : {}),
+            },
+        }, {
+            apiVersion: fields.apiVersion,
+            baseUrl: fields.baseUrl,
+            customHeaders: fields.customHeaders,
+        });
+        this.streamUsage = fields.streamUsage ?? this.streamUsage;
+    }
+    static lc_name() {
+        return 'IllumaGoogleGenerativeAI';
+    }
+    /**
+     * Helper function to convert Gemini API usage metadata to LangChain format
+     * Includes support for cached tokens and tier-based tracking for gemini-3-pro-preview
+     */
+    _convertToUsageMetadata(usageMetadata, model) {
+        if (!usageMetadata) {
+            return undefined;
+        }
+        const output = {
+            input_tokens: usageMetadata.promptTokenCount ?? 0,
+            output_tokens: (usageMetadata.candidatesTokenCount ?? 0) +
+                (usageMetadata.thoughtsTokenCount ?? 0),
+            total_tokens: usageMetadata.totalTokenCount ?? 0,
+        };
+        if (usageMetadata.cachedContentTokenCount) {
+            output.input_token_details ??= {};
+            output.input_token_details.cache_read =
+                usageMetadata.cachedContentTokenCount;
+        }
+        // gemini-3-pro-preview has bracket based tracking of tokens per request
+        if (model === 'gemini-3-pro-preview') {
+            const over200k = Math.max(0, (usageMetadata.promptTokenCount ?? 0) - 200000);
+            const cachedOver200k = Math.max(0, (usageMetadata.cachedContentTokenCount ?? 0) - 200000);
+            if (over200k) {
+                output.input_token_details = {
+                    ...output.input_token_details,
+                    over_200k: over200k,
+                };
+            }
+            if (cachedOver200k) {
+                output.input_token_details = {
+                    ...output.input_token_details,
+                    cache_read_over_200k: cachedOver200k,
+                };
+            }
+        }
+        return output;
+    }
+    invocationParams(options) {
+        const params = super.invocationParams(options);
+        if (this.thinkingConfig) {
+            /** @ts-ignore */
+            this.client.generationConfig = {
+                /** @ts-ignore */
+                ...this.client.generationConfig,
+                /** @ts-ignore */
+                thinkingConfig: this.thinkingConfig,
+            };
+        }
+        return params;
+    }
+    async _generate(messages, options, runManager) {
+        const prompt = common.convertBaseMessagesToContent(messages, this._isMultimodalModel, this.useSystemInstruction, this.model);
+        let actualPrompt = prompt;
+        if (prompt?.[0].role === 'system') {
+            const [systemInstruction] = prompt;
+            /** @ts-ignore */
+            this.client.systemInstruction = systemInstruction;
+            actualPrompt = prompt.slice(1);
+        }
+        const parameters = this.invocationParams(options);
+        const request = {
+            ...parameters,
+            contents: actualPrompt,
+        };
+        const res = await this.caller.callWithOptions({ signal: options.signal }, async () => 
+        /** @ts-ignore */
+        this.client.generateContent(request));
+        const response = res.response;
+        const usageMetadata = this._convertToUsageMetadata(
+        /** @ts-ignore */
+        response.usageMetadata, this.model);
+        /** @ts-ignore */
+        const generationResult = common.mapGenerateContentResultToChatResult(response, {
+            usageMetadata,
+        });
+        await runManager?.handleLLMNewToken(generationResult.generations[0].text || '', undefined, undefined, undefined, undefined, undefined);
+        return generationResult;
+    }
+    async *_streamResponseChunks(messages$1, options, runManager) {
+        const prompt = common.convertBaseMessagesToContent(messages$1, this._isMultimodalModel, this.useSystemInstruction, this.model);
+        let actualPrompt = prompt;
+        if (prompt?.[0].role === 'system') {
+            const [systemInstruction] = prompt;
+            /** @ts-ignore */
+            this.client.systemInstruction = systemInstruction;
+            actualPrompt = prompt.slice(1);
+        }
+        const parameters = this.invocationParams(options);
+        const request = {
+            ...parameters,
+            contents: actualPrompt,
+        };
+        const stream = await this.caller.callWithOptions({ signal: options.signal }, async () => {
+            /** @ts-ignore */
+            const { stream } = await this.client.generateContentStream(request);
+            return stream;
+        });
+        let lastUsageMetadata;
+        for await (const response of stream) {
+            if ('usageMetadata' in response &&
+                this.streamUsage !== false &&
+                options.streamUsage !== false) {
+                lastUsageMetadata = this._convertToUsageMetadata(response.usageMetadata, this.model);
+            }
+            const chunk = common.convertResponseContentToChatGenerationChunk(response, {
+                usageMetadata: undefined});
+            if (!chunk) {
+                continue;
+            }
+            yield chunk;
+            await runManager?.handleLLMNewToken(chunk.text || '', undefined, undefined, undefined, undefined, { chunk });
+        }
+        if (lastUsageMetadata) {
+            const finalChunk = new outputs.ChatGenerationChunk({
+                text: '',
+                message: new messages.AIMessageChunk({
+                    content: '',
+                    usage_metadata: lastUsageMetadata,
+                }),
+            });
+            yield finalChunk;
+            await runManager?.handleLLMNewToken(finalChunk.text || '', undefined, undefined, undefined, undefined, { chunk: finalChunk });
+        }
+    }
+}
+
+exports.CustomChatGoogleGenerativeAI = CustomChatGoogleGenerativeAI;
+//# sourceMappingURL=index.cjs.map