illuma-agents 1.0.53 → 1.0.54

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

@@ -1 +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\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;;AAGxC,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;;;;"}
+{"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/esm/graphs/Graph.mjs

@@ -482,26 +482,45 @@ class StandardGraph extends Graph {
             method = undefined; // Let LangChain choose the default
         }
         // Use withStructuredOutput to bind the schema
+        // Always use includeRaw: true internally so we can debug what's returned
         // eslint-disable-next-line @typescript-eslint/no-explicit-any
         const structuredModel = model.withStructuredOutput(schema, {
             name,
             method,
-            includeRaw,
+            includeRaw: true, // Always true internally for debugging
             strict: structuredOutputConfig.strict !== false,
         });
+        console.log('[Graph] Structured output config:', {
+            name,
+            method,
+            provider,
+            schemaKeys: Object.keys(schema),
+            // eslint-disable-next-line @typescript-eslint/no-explicit-any
+            modelName: model.model || model.modelId || 'unknown',
+        });
         let lastError;
         let attempts = 0;
         while (attempts <= maxRetries) {
             try {
                 const result = await structuredModel.invoke(finalMessages, config);
-                // Handle includeRaw response format
-                if (includeRaw && result.raw && result.parsed) {
+                // Debug: log what we got back
+                console.log('[Graph] Structured output raw result type:', typeof result);
+                if (result?.raw) {
+                    const rawMsg = result.raw;
+                    console.log('[Graph] Raw message content type:', typeof rawMsg?.content);
+                    console.log('[Graph] Raw message tool_calls:', rawMsg?.tool_calls?.length ?? 0);
+                    if (rawMsg?.content && typeof rawMsg.content === 'string' && rawMsg.content.length > 0) {
+                        console.log('[Graph] Raw message text content (first 200):', rawMsg.content.substring(0, 200));
+                    }
+                }
+                // Handle response - we always use includeRaw internally
+                if (result?.raw && result?.parsed !== undefined) {
                     return {
                         structuredResponse: result.parsed,
                         rawMessage: result.raw,
                     };
                 }
-                // Direct response
+                // Fallback for models that don't support includeRaw
                 return {
                     structuredResponse: result,
                 };