langchain 1.0.0-alpha.7 → 1.0.0-alpha.9

package/dist/agents/middleware/promptCaching.d.ts

@@ -0,0 +1,197 @@
+import { AgentMiddleware } from "./types.js";
+import { z } from "zod/v3";
+import { InferInteropZodInput } from "@langchain/core/utils/types";
+
+//#region src/agents/middleware/promptCaching.d.ts
+declare const contextSchema: z.ZodObject<{
+  /**
+   * Whether to enable prompt caching.
+   * @default true
+   */
+  enableCaching: z.ZodOptional<z.ZodBoolean>;
+  /**
+   * The time-to-live for the cached prompt.
+   * @default "5m"
+   */
+  ttl: z.ZodOptional<z.ZodEnum<["5m", "1h"]>>;
+  /**
+   * The minimum number of messages required before caching is applied.
+   * @default 3
+   */
+  minMessagesToCache: z.ZodOptional<z.ZodNumber>;
+  /**
+   * The behavior to take when an unsupported model is used.
+   * - "ignore" will ignore the unsupported model and continue without caching.
+   * - "warn" will warn the user and continue without caching.
+   * - "raise" will raise an error and stop the agent.
+   * @default "warn"
+   */
+  unsupportedModelBehavior: z.ZodOptional<z.ZodEnum<["ignore", "warn", "raise"]>>;
+}, "strip", z.ZodTypeAny, {
+  enableCaching?: boolean | undefined;
+  ttl?: "1h" | "5m" | undefined;
+  minMessagesToCache?: number | undefined;
+  unsupportedModelBehavior?: "ignore" | "raise" | "warn" | undefined;
+}, {
+  enableCaching?: boolean | undefined;
+  ttl?: "1h" | "5m" | undefined;
+  minMessagesToCache?: number | undefined;
+  unsupportedModelBehavior?: "ignore" | "raise" | "warn" | undefined;
+}>;
+type PromptCachingMiddlewareConfig = Partial<InferInteropZodInput<typeof contextSchema>>;
+/**
+ * Creates a prompt caching middleware for Anthropic models to optimize API usage.
+ *
+ * This middleware automatically adds cache control headers to the last messages when using Anthropic models,
+ * enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive
+ * prompts, long system messages, or extensive conversation histories.
+ *
+ * ## How It Works
+ *
+ * The middleware intercepts model requests and adds cache control metadata that tells Anthropic's
+ * API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the
+ * cached representations are reused, skipping redundant token processing.
+ *
+ * ## Benefits
+ *
+ * - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)
+ * - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed
+ * - **Better Scalability**: Reduced computational load enables handling more requests
+ * - **Consistent Performance**: Stable response times for repetitive queries
+ *
+ * @param middlewareOptions - Configuration options for the caching behavior
+ * @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)
+ * @param middlewareOptions.ttl - Cache time-to-live: `"5m"` for 5 minutes or `"1h"` for 1 hour (default: `"5m"`)
+ * @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)
+ * @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `"warn"`)
+ *
+ * @returns A middleware instance that can be passed to `createAgent`
+ *
+ * @throws {Error} If used with non-Anthropic models
+ *
+ * @example
+ * Basic usage with default settings
+ * ```typescript
+ * import { createAgent } from "langchain";
+ * import { anthropicPromptCachingMiddleware } from "langchain";
+ *
+ * const agent = createAgent({
+ *   model: "anthropic:claude-3-5-sonnet",
+ *   middleware: [
+ *     anthropicPromptCachingMiddleware()
+ *   ]
+ * });
+ * ```
+ *
+ * @example
+ * Custom configuration for longer conversations
+ * ```typescript
+ * const cachingMiddleware = anthropicPromptCachingMiddleware({
+ *   ttl: "1h",  // Cache for 1 hour instead of default 5 minutes
+ *   minMessagesToCache: 5  // Only cache after 5 messages
+ * });
+ *
+ * const agent = createAgent({
+ *   model: "anthropic:claude-3-5-sonnet",
+ *   systemPrompt: "You are a helpful assistant with deep knowledge of...", // Long system prompt
+ *   middleware: [cachingMiddleware]
+ * });
+ * ```
+ *
+ * @example
+ * Conditional caching based on runtime context
+ * ```typescript
+ * const agent = createAgent({
+ *   model: "anthropic:claude-3-5-sonnet",
+ *   middleware: [
+ *     anthropicPromptCachingMiddleware({
+ *       enableCaching: true,
+ *       ttl: "5m"
+ *     })
+ *   ]
+ * });
+ *
+ * // Disable caching for specific requests
+ * await agent.invoke(
+ *   { messages: [new HumanMessage("Process this without caching")] },
+ *   {
+ *     configurable: {
+ *       middleware_context: { enableCaching: false }
+ *     }
+ *   }
+ * );
+ * ```
+ *
+ * @example
+ * Optimal setup for customer support chatbot
+ * ```typescript
+ * const supportAgent = createAgent({
+ *   model: "anthropic:claude-3-5-sonnet",
+ *   systemPrompt: `You are a customer support agent for ACME Corp.
+ *
+ *     Company policies:
+ *     - Always be polite and professional
+ *     - Refer to knowledge base for product information
+ *     - Escalate billing issues to human agents
+ *     ... (extensive policies and guidelines)
+ *   `,
+ *   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],
+ *   middleware: [
+ *     anthropicPromptCachingMiddleware({
+ *       ttl: "1h",  // Long TTL for stable system prompt
+ *       minMessagesToCache: 1  // Cache immediately due to large system prompt
+ *     })
+ *   ]
+ * });
+ * ```
+ *
+ * @remarks
+ * - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers
+ * - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`
+ * - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys
+ * - **TTL Options**: Only supports "5m" (5 minutes) and "1h" (1 hour) as TTL values per Anthropic's API
+ * - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications
+ * - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base
+ *
+ * @see {@link createAgent} for agent creation
+ * @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation
+ * @public
+ */
+declare function anthropicPromptCachingMiddleware(middlewareOptions?: PromptCachingMiddlewareConfig): AgentMiddleware<undefined, z.ZodObject<{
+  /**
+   * Whether to enable prompt caching.
+   * @default true
+   */
+  enableCaching: z.ZodOptional<z.ZodBoolean>;
+  /**
+   * The time-to-live for the cached prompt.
+   * @default "5m"
+   */
+  ttl: z.ZodOptional<z.ZodEnum<["5m", "1h"]>>;
+  /**
+   * The minimum number of messages required before caching is applied.
+   * @default 3
+   */
+  minMessagesToCache: z.ZodOptional<z.ZodNumber>;
+  /**
+   * The behavior to take when an unsupported model is used.
+   * - "ignore" will ignore the unsupported model and continue without caching.
+   * - "warn" will warn the user and continue without caching.
+   * - "raise" will raise an error and stop the agent.
+   * @default "warn"
+   */
+  unsupportedModelBehavior: z.ZodOptional<z.ZodEnum<["ignore", "warn", "raise"]>>;
+}, "strip", z.ZodTypeAny, {
+  enableCaching?: boolean | undefined;
+  ttl?: "1h" | "5m" | undefined;
+  minMessagesToCache?: number | undefined;
+  unsupportedModelBehavior?: "ignore" | "raise" | "warn" | undefined;
+}, {
+  enableCaching?: boolean | undefined;
+  ttl?: "1h" | "5m" | undefined;
+  minMessagesToCache?: number | undefined;
+  unsupportedModelBehavior?: "ignore" | "raise" | "warn" | undefined;
+}>, any>;
+//#endregion
+export { PromptCachingMiddlewareConfig, anthropicPromptCachingMiddleware };
+//# sourceMappingURL=promptCaching.d.ts.map

package/dist/agents/middleware/promptCaching.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"promptCaching.d.ts","names":["z","InferInteropZodInput","contextSchema","ZodBoolean","ZodOptional","ZodEnum","ZodNumber","ZodTypeAny","ZodObject","PromptCachingMiddlewareConfig","Partial","anthropicPromptCachingMiddleware","__types_js1","AgentMiddleware"],"sources":["../../../src/agents/middleware/promptCaching.d.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { InferInteropZodInput } from \"@langchain/core/utils/types\";\ndeclare const contextSchema: z.ZodObject<{\n    /**\n     * Whether to enable prompt caching.\n     * @default true\n     */\n    enableCaching: z.ZodOptional<z.ZodBoolean>;\n    /**\n     * The time-to-live for the cached prompt.\n     * @default \"5m\"\n     */\n    ttl: z.ZodOptional<z.ZodEnum<[\"5m\", \"1h\"]>>;\n    /**\n     * The minimum number of messages required before caching is applied.\n     * @default 3\n     */\n    minMessagesToCache: z.ZodOptional<z.ZodNumber>;\n    /**\n     * The behavior to take when an unsupported model is used.\n     * - \"ignore\" will ignore the unsupported model and continue without caching.\n     * - \"warn\" will warn the user and continue without caching.\n     * - \"raise\" will raise an error and stop the agent.\n     * @default \"warn\"\n     */\n    unsupportedModelBehavior: z.ZodOptional<z.ZodEnum<[\"ignore\", \"warn\", \"raise\"]>>;\n}, \"strip\", z.ZodTypeAny, {\n    enableCaching?: boolean | undefined;\n    ttl?: \"1h\" | \"5m\" | undefined;\n    minMessagesToCache?: number | undefined;\n    unsupportedModelBehavior?: \"ignore\" | \"raise\" | \"warn\" | undefined;\n}, {\n    enableCaching?: boolean | undefined;\n    ttl?: \"1h\" | \"5m\" | undefined;\n    minMessagesToCache?: number | undefined;\n    unsupportedModelBehavior?: \"ignore\" | \"raise\" | \"warn\" | undefined;\n}>;\nexport type PromptCachingMiddlewareConfig = Partial<InferInteropZodInput<typeof contextSchema>>;\n/**\n * Creates a prompt caching middleware for Anthropic models to optimize API usage.\n *\n * This middleware automatically adds cache control headers to the last messages when using Anthropic models,\n * enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive\n * prompts, long system messages, or extensive conversation histories.\n *\n * ## How It Works\n *\n * The middleware intercepts model requests and adds cache control metadata that tells Anthropic's\n * API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the\n * cached representations are reused, skipping redundant token processing.\n *\n * ## Benefits\n *\n * - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)\n * - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed\n * - **Better Scalability**: Reduced computational load enables handling more requests\n * - **Consistent Performance**: Stable response times for repetitive queries\n *\n * @param middlewareOptions - Configuration options for the caching behavior\n * @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)\n * @param middlewareOptions.ttl - Cache time-to-live: `\"5m\"` for 5 minutes or `\"1h\"` for 1 hour (default: `\"5m\"`)\n * @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)\n * @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `\"warn\"`)\n *\n * @returns A middleware instance that can be passed to `createAgent`\n *\n * @throws {Error} If used with non-Anthropic models\n *\n * @example\n * Basic usage with default settings\n * ```typescript\n * import { createAgent } from \"langchain\";\n * import { anthropicPromptCachingMiddleware } from \"langchain\";\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware()\n *   ]\n * });\n * ```\n *\n * @example\n * Custom configuration for longer conversations\n * ```typescript\n * const cachingMiddleware = anthropicPromptCachingMiddleware({\n *   ttl: \"1h\",  // Cache for 1 hour instead of default 5 minutes\n *   minMessagesToCache: 5  // Only cache after 5 messages\n * });\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: \"You are a helpful assistant with deep knowledge of...\", // Long system prompt\n *   middleware: [cachingMiddleware]\n * });\n * ```\n *\n * @example\n * Conditional caching based on runtime context\n * ```typescript\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       enableCaching: true,\n *       ttl: \"5m\"\n *     })\n *   ]\n * });\n *\n * // Disable caching for specific requests\n * await agent.invoke(\n *   { messages: [new HumanMessage(\"Process this without caching\")] },\n *   {\n *     configurable: {\n *       middleware_context: { enableCaching: false }\n *     }\n *   }\n * );\n * ```\n *\n * @example\n * Optimal setup for customer support chatbot\n * ```typescript\n * const supportAgent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: `You are a customer support agent for ACME Corp.\n *\n *     Company policies:\n *     - Always be polite and professional\n *     - Refer to knowledge base for product information\n *     - Escalate billing issues to human agents\n *     ... (extensive policies and guidelines)\n *   `,\n *   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       ttl: \"1h\",  // Long TTL for stable system prompt\n *       minMessagesToCache: 1  // Cache immediately due to large system prompt\n *     })\n *   ]\n * });\n * ```\n *\n * @remarks\n * - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers\n * - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`\n * - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys\n * - **TTL Options**: Only supports \"5m\" (5 minutes) and \"1h\" (1 hour) as TTL values per Anthropic's API\n * - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications\n * - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base\n *\n * @see {@link createAgent} for agent creation\n * @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation\n * @public\n */\nexport declare function anthropicPromptCachingMiddleware(middlewareOptions?: PromptCachingMiddlewareConfig): import(\"./types.js\").AgentMiddleware<undefined, z.ZodObject<{\n    /**\n     * Whether to enable prompt caching.\n     * @default true\n     */\n    enableCaching: z.ZodOptional<z.ZodBoolean>;\n    /**\n     * The time-to-live for the cached prompt.\n     * @default \"5m\"\n     */\n    ttl: z.ZodOptional<z.ZodEnum<[\"5m\", \"1h\"]>>;\n    /**\n     * The minimum number of messages required before caching is applied.\n     * @default 3\n     */\n    minMessagesToCache: z.ZodOptional<z.ZodNumber>;\n    /**\n     * The behavior to take when an unsupported model is used.\n     * - \"ignore\" will ignore the unsupported model and continue without caching.\n     * - \"warn\" will warn the user and continue without caching.\n     * - \"raise\" will raise an error and stop the agent.\n     * @default \"warn\"\n     */\n    unsupportedModelBehavior: z.ZodOptional<z.ZodEnum<[\"ignore\", \"warn\", \"raise\"]>>;\n}, \"strip\", z.ZodTypeAny, {\n    enableCaching?: boolean | undefined;\n    ttl?: \"1h\" | \"5m\" | undefined;\n    minMessagesToCache?: number | undefined;\n    unsupportedModelBehavior?: \"ignore\" | \"raise\" | \"warn\" | undefined;\n}, {\n    enableCaching?: boolean | undefined;\n    ttl?: \"1h\" | \"5m\" | undefined;\n    minMessagesToCache?: number | undefined;\n    unsupportedModelBehavior?: \"ignore\" | \"raise\" | \"warn\" | undefined;\n}>, any>;\nexport {};\n"],"mappings":";;;;;cAEcE,eAAeF,CAAAA,CAAEQ;;;AADoC;;EAmCjE,aA7BiCL,EAAhBH,CAAAA,CAAEI,WAAcD,CAAFH,CAAAA,CAAEG,UAAAA,CAAAA;EAAU;;;;EAUI,GAAzBH,EALfA,CAAAA,CAAEI,WAKeA,CALHJ,CAAAA,CAAEK,OAKCD,CAAAA,CAAAA,IAAAA,EAAAA,IAAAA,CAAAA,CAAAA,CAAAA;EAAW;;;;EAfG,kBAAA,EAehBJ,CAAAA,CAAEI,WAfc,CAeFJ,CAAAA,CAAEM,SAfA,CAAA;EAmC5BG;;;;;AAAuC;AAuHnD;EAAwD,wBAAA,EAnI1BT,CAAAA,CAAEI,WAmIwB,CAnIZJ,CAAAA,CAAEK,OAmIU,CAAA,CAAA,QAAA,EAAA,MAAA,EAAA,OAAA,CAAA,CAAA,CAAA;CAAA,EAAA,OAAqBI,EAlIjET,CAAAA,CAAEO,UAkI+DE,EAAAA;EAA6B,aAKvEN,CAAAA,EAAAA,OAAAA,GAAAA,SAAAA;EAAU,GAA1BH,CAAAA,EAAEI,IAAAA,GAAAA,IAAAA,GAAAA,SAAAA;EAAW,kBAKPC,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAO,wBAArBD,CAAAA,EAAAA,QAAAA,GAAAA,OAAAA,GAAAA,MAAAA,GAAAA,SAAAA;CAAW,EAAA;EAK2B,aAAvBA,CAAAA,EAAAA,OAAAA,GAAAA,SAAAA;EAAW,GAQOJ,CAAAA,EAAEK,IAAAA,GAAAA,IAAAA,GAAAA,SAAAA;EAAO,kBAArBD,CAAAA,EAAAA,MAAAA,GAAAA,SAAAA;EAAW,wBAC7BG,CAAAA,EAAAA,QAAAA,GAAAA,OAAAA,GAAAA,MAAAA,GAAAA,SAAAA;CAAU,CAAA;AAxBgJ,KAvH5JE,6BAAAA,GAAgCC,OAuH4H,CAvHpHT,oBAuHoH,CAAA,OAvHxFC,aAuHwF,CAAA,CAAA;AAAvB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBAAzHS,gCAAAA,qBAAqDF,gCAA2F,2BAAXT,CAAAA,CAAEQ;;;;;iBAK5IR,CAAAA,CAAEI,YAAYJ,CAAAA,CAAEG;;;;;OAK1BH,CAAAA,CAAEI,YAAYJ,CAAAA,CAAEK;;;;;sBAKDL,CAAAA,CAAEI,YAAYJ,CAAAA,CAAEM;;;;;;;;4BAQVN,CAAAA,CAAEI,YAAYJ,CAAAA,CAAEK;YAClCL,CAAAA,CAAEO"}

package/dist/agents/middleware/promptCaching.js

@@ -0,0 +1,212 @@
+import { createMiddleware } from "../middleware.js";
+import { z } from "zod/v3";
+
+//#region src/agents/middleware/promptCaching.ts
+const DEFAULT_ENABLE_CACHING = true;
+const DEFAULT_TTL = "5m";
+const DEFAULT_MIN_MESSAGES_TO_CACHE = 3;
+const DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR = "warn";
+const contextSchema = z.object({
+	enableCaching: z.boolean().optional(),
+	ttl: z.enum(["5m", "1h"]).optional(),
+	minMessagesToCache: z.number().optional(),
+	unsupportedModelBehavior: z.enum([
+		"ignore",
+		"warn",
+		"raise"
+	]).optional()
+});
+var PromptCachingMiddlewareError = class extends Error {
+	constructor(message) {
+		super(message);
+		this.name = "PromptCachingMiddlewareError";
+	}
+};
+/**
+* Creates a prompt caching middleware for Anthropic models to optimize API usage.
+*
+* This middleware automatically adds cache control headers to the last messages when using Anthropic models,
+* enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive
+* prompts, long system messages, or extensive conversation histories.
+*
+* ## How It Works
+*
+* The middleware intercepts model requests and adds cache control metadata that tells Anthropic's
+* API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the
+* cached representations are reused, skipping redundant token processing.
+*
+* ## Benefits
+*
+* - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)
+* - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed
+* - **Better Scalability**: Reduced computational load enables handling more requests
+* - **Consistent Performance**: Stable response times for repetitive queries
+*
+* @param middlewareOptions - Configuration options for the caching behavior
+* @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)
+* @param middlewareOptions.ttl - Cache time-to-live: `"5m"` for 5 minutes or `"1h"` for 1 hour (default: `"5m"`)
+* @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)
+* @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `"warn"`)
+*
+* @returns A middleware instance that can be passed to `createAgent`
+*
+* @throws {Error} If used with non-Anthropic models
+*
+* @example
+* Basic usage with default settings
+* ```typescript
+* import { createAgent } from "langchain";
+* import { anthropicPromptCachingMiddleware } from "langchain";
+*
+* const agent = createAgent({
+*   model: "anthropic:claude-3-5-sonnet",
+*   middleware: [
+*     anthropicPromptCachingMiddleware()
+*   ]
+* });
+* ```
+*
+* @example
+* Custom configuration for longer conversations
+* ```typescript
+* const cachingMiddleware = anthropicPromptCachingMiddleware({
+*   ttl: "1h",  // Cache for 1 hour instead of default 5 minutes
+*   minMessagesToCache: 5  // Only cache after 5 messages
+* });
+*
+* const agent = createAgent({
+*   model: "anthropic:claude-3-5-sonnet",
+*   systemPrompt: "You are a helpful assistant with deep knowledge of...", // Long system prompt
+*   middleware: [cachingMiddleware]
+* });
+* ```
+*
+* @example
+* Conditional caching based on runtime context
+* ```typescript
+* const agent = createAgent({
+*   model: "anthropic:claude-3-5-sonnet",
+*   middleware: [
+*     anthropicPromptCachingMiddleware({
+*       enableCaching: true,
+*       ttl: "5m"
+*     })
+*   ]
+* });
+*
+* // Disable caching for specific requests
+* await agent.invoke(
+*   { messages: [new HumanMessage("Process this without caching")] },
+*   {
+*     configurable: {
+*       middleware_context: { enableCaching: false }
+*     }
+*   }
+* );
+* ```
+*
+* @example
+* Optimal setup for customer support chatbot
+* ```typescript
+* const supportAgent = createAgent({
+*   model: "anthropic:claude-3-5-sonnet",
+*   systemPrompt: `You are a customer support agent for ACME Corp.
+*
+*     Company policies:
+*     - Always be polite and professional
+*     - Refer to knowledge base for product information
+*     - Escalate billing issues to human agents
+*     ... (extensive policies and guidelines)
+*   `,
+*   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],
+*   middleware: [
+*     anthropicPromptCachingMiddleware({
+*       ttl: "1h",  // Long TTL for stable system prompt
+*       minMessagesToCache: 1  // Cache immediately due to large system prompt
+*     })
+*   ]
+* });
+* ```
+*
+* @remarks
+* - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers
+* - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`
+* - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys
+* - **TTL Options**: Only supports "5m" (5 minutes) and "1h" (1 hour) as TTL values per Anthropic's API
+* - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications
+* - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base
+*
+* @see {@link createAgent} for agent creation
+* @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation
+* @public
+*/
+function anthropicPromptCachingMiddleware(middlewareOptions) {
+	return createMiddleware({
+		name: "PromptCachingMiddleware",
+		contextSchema,
+		wrapModelCall: (request, handler) => {
+			/**
+			* Prefer runtime context values over middleware options values over defaults
+			*/
+			const enableCaching = request.runtime.context.enableCaching ?? middlewareOptions?.enableCaching ?? DEFAULT_ENABLE_CACHING;
+			const ttl = request.runtime.context.ttl ?? middlewareOptions?.ttl ?? DEFAULT_TTL;
+			const minMessagesToCache = request.runtime.context.minMessagesToCache ?? middlewareOptions?.minMessagesToCache ?? DEFAULT_MIN_MESSAGES_TO_CACHE;
+			const unsupportedModelBehavior = request.runtime.context.unsupportedModelBehavior ?? middlewareOptions?.unsupportedModelBehavior ?? DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR;
+			if (!enableCaching || !request.model) return handler(request);
+			const isAnthropicModel = request.model.getName() === "ChatAnthropic" || request.model.getName() === "ConfigurableModel" && request.model._defaultConfig?.modelProvider === "anthropic";
+			if (!isAnthropicModel) {
+				const modelName = request.model.getName();
+				const modelInfo = request.model.getName() === "ConfigurableModel" ? `${modelName} (${request.model._defaultConfig?.modelProvider})` : modelName;
+				const baseMessage = `Unsupported model '${modelInfo}'. Prompt caching requires an Anthropic model`;
+				if (unsupportedModelBehavior === "raise") throw new PromptCachingMiddlewareError(`${baseMessage} (e.g., 'anthropic:claude-4-0-sonnet').`);
+				else if (unsupportedModelBehavior === "warn") console.warn(`PromptCachingMiddleware: Skipping caching for ${modelName}. Consider switching to an Anthropic model for caching benefits.`);
+				return handler(request);
+			}
+			const messagesCount = request.state.messages.length + (request.systemPrompt ? 1 : 0);
+			if (messagesCount < minMessagesToCache) return handler(request);
+			/**
+			* Add cache_control to the last message
+			*/
+			const lastMessage = request.messages.at(-1);
+			if (!lastMessage) return handler(request);
+			const NewMessageConstructor = Object.getPrototypeOf(lastMessage).constructor;
+			if (Array.isArray(lastMessage.content)) {
+				const newMessage = new NewMessageConstructor({
+					...lastMessage,
+					content: [...lastMessage.content.slice(0, -1), {
+						...lastMessage.content.at(-1),
+						cache_control: {
+							type: "ephemeral",
+							ttl
+						}
+					}]
+				});
+				return handler({
+					...request,
+					messages: [...request.messages.slice(0, -1), newMessage]
+				});
+			} else if (typeof lastMessage.content === "string") {
+				const newMessage = new NewMessageConstructor({
+					...lastMessage,
+					content: [{
+						type: "text",
+						text: lastMessage.content,
+						cache_control: {
+							type: "ephemeral",
+							ttl
+						}
+					}]
+				});
+				return handler({
+					...request,
+					messages: [...request.messages.slice(0, -1), newMessage]
+				});
+			}
+			throw new PromptCachingMiddlewareError("Last message content is not a string or array");
+		}
+	});
+}
+
+//#endregion
+export { anthropicPromptCachingMiddleware };
+//# sourceMappingURL=promptCaching.js.map

package/dist/agents/middleware/promptCaching.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"promptCaching.js","names":["message: string","middlewareOptions?: PromptCachingMiddlewareConfig"],"sources":["../../../src/agents/middleware/promptCaching.ts"],"sourcesContent":["import { z } from \"zod/v3\";\nimport { ContentBlock } from \"@langchain/core/messages\";\nimport { InferInteropZodInput } from \"@langchain/core/utils/types\";\n\nimport { ConfigurableModel } from \"../../chat_models/universal.js\";\nimport { createMiddleware } from \"../middleware.js\";\n\nconst DEFAULT_ENABLE_CACHING = true;\nconst DEFAULT_TTL = \"5m\";\nconst DEFAULT_MIN_MESSAGES_TO_CACHE = 3;\nconst DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR = \"warn\";\n\nconst contextSchema = z.object({\n  /**\n   * Whether to enable prompt caching.\n   * @default true\n   */\n  enableCaching: z.boolean().optional(),\n  /**\n   * The time-to-live for the cached prompt.\n   * @default \"5m\"\n   */\n  ttl: z.enum([\"5m\", \"1h\"]).optional(),\n  /**\n   * The minimum number of messages required before caching is applied.\n   * @default 3\n   */\n  minMessagesToCache: z.number().optional(),\n  /**\n   * The behavior to take when an unsupported model is used.\n   * - \"ignore\" will ignore the unsupported model and continue without caching.\n   * - \"warn\" will warn the user and continue without caching.\n   * - \"raise\" will raise an error and stop the agent.\n   * @default \"warn\"\n   */\n  unsupportedModelBehavior: z.enum([\"ignore\", \"warn\", \"raise\"]).optional(),\n});\nexport type PromptCachingMiddlewareConfig = Partial<\n  InferInteropZodInput<typeof contextSchema>\n>;\n\nclass PromptCachingMiddlewareError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"PromptCachingMiddlewareError\";\n  }\n}\n\n/**\n * Creates a prompt caching middleware for Anthropic models to optimize API usage.\n *\n * This middleware automatically adds cache control headers to the last messages when using Anthropic models,\n * enabling their prompt caching feature. This can significantly reduce costs for applications with repetitive\n * prompts, long system messages, or extensive conversation histories.\n *\n * ## How It Works\n *\n * The middleware intercepts model requests and adds cache control metadata that tells Anthropic's\n * API to cache processed prompt prefixes. On subsequent requests with matching prefixes, the\n * cached representations are reused, skipping redundant token processing.\n *\n * ## Benefits\n *\n * - **Cost Reduction**: Avoid reprocessing the same tokens repeatedly (up to 90% savings on cached portions)\n * - **Lower Latency**: Cached prompts are processed faster as embeddings are pre-computed\n * - **Better Scalability**: Reduced computational load enables handling more requests\n * - **Consistent Performance**: Stable response times for repetitive queries\n *\n * @param middlewareOptions - Configuration options for the caching behavior\n * @param middlewareOptions.enableCaching - Whether to enable prompt caching (default: `true`)\n * @param middlewareOptions.ttl - Cache time-to-live: `\"5m\"` for 5 minutes or `\"1h\"` for 1 hour (default: `\"5m\"`)\n * @param middlewareOptions.minMessagesToCache - Minimum number of messages required before caching is applied (default: `3`)\n * @param middlewareOptions.unsupportedModelBehavior - The behavior to take when an unsupported model is used (default: `\"warn\"`)\n *\n * @returns A middleware instance that can be passed to `createAgent`\n *\n * @throws {Error} If used with non-Anthropic models\n *\n * @example\n * Basic usage with default settings\n * ```typescript\n * import { createAgent } from \"langchain\";\n * import { anthropicPromptCachingMiddleware } from \"langchain\";\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware()\n *   ]\n * });\n * ```\n *\n * @example\n * Custom configuration for longer conversations\n * ```typescript\n * const cachingMiddleware = anthropicPromptCachingMiddleware({\n *   ttl: \"1h\",  // Cache for 1 hour instead of default 5 minutes\n *   minMessagesToCache: 5  // Only cache after 5 messages\n * });\n *\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: \"You are a helpful assistant with deep knowledge of...\", // Long system prompt\n *   middleware: [cachingMiddleware]\n * });\n * ```\n *\n * @example\n * Conditional caching based on runtime context\n * ```typescript\n * const agent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       enableCaching: true,\n *       ttl: \"5m\"\n *     })\n *   ]\n * });\n *\n * // Disable caching for specific requests\n * await agent.invoke(\n *   { messages: [new HumanMessage(\"Process this without caching\")] },\n *   {\n *     configurable: {\n *       middleware_context: { enableCaching: false }\n *     }\n *   }\n * );\n * ```\n *\n * @example\n * Optimal setup for customer support chatbot\n * ```typescript\n * const supportAgent = createAgent({\n *   model: \"anthropic:claude-3-5-sonnet\",\n *   systemPrompt: `You are a customer support agent for ACME Corp.\n *\n *     Company policies:\n *     - Always be polite and professional\n *     - Refer to knowledge base for product information\n *     - Escalate billing issues to human agents\n *     ... (extensive policies and guidelines)\n *   `,\n *   tools: [searchKnowledgeBase, createTicket, checkOrderStatus],\n *   middleware: [\n *     anthropicPromptCachingMiddleware({\n *       ttl: \"1h\",  // Long TTL for stable system prompt\n *       minMessagesToCache: 1  // Cache immediately due to large system prompt\n *     })\n *   ]\n * });\n * ```\n *\n * @remarks\n * - **Anthropic Only**: This middleware only works with Anthropic models and will throw an error if used with other providers\n * - **Automatic Application**: Caching is applied automatically when message count exceeds `minMessagesToCache`\n * - **Cache Scope**: Caches are isolated per API key and cannot be shared across different keys\n * - **TTL Options**: Only supports \"5m\" (5 minutes) and \"1h\" (1 hour) as TTL values per Anthropic's API\n * - **Best Use Cases**: Long system prompts, multi-turn conversations, repetitive queries, RAG applications\n * - **Cost Impact**: Cached tokens are billed at 10% of the base input token price, cache writes are billed at 25% of the base\n *\n * @see {@link createAgent} for agent creation\n * @see {@link https://docs.anthropic.com/en/docs/build-with-claude/prompt-caching} Anthropic's prompt caching documentation\n * @public\n */\nexport function anthropicPromptCachingMiddleware(\n  middlewareOptions?: PromptCachingMiddlewareConfig\n) {\n  return createMiddleware({\n    name: \"PromptCachingMiddleware\",\n    contextSchema,\n    wrapModelCall: (request, handler) => {\n      /**\n       * Prefer runtime context values over middleware options values over defaults\n       */\n      const enableCaching =\n        request.runtime.context.enableCaching ??\n        middlewareOptions?.enableCaching ??\n        DEFAULT_ENABLE_CACHING;\n      const ttl =\n        request.runtime.context.ttl ?? middlewareOptions?.ttl ?? DEFAULT_TTL;\n      const minMessagesToCache =\n        request.runtime.context.minMessagesToCache ??\n        middlewareOptions?.minMessagesToCache ??\n        DEFAULT_MIN_MESSAGES_TO_CACHE;\n      const unsupportedModelBehavior =\n        request.runtime.context.unsupportedModelBehavior ??\n        middlewareOptions?.unsupportedModelBehavior ??\n        DEFAULT_UNSUPPORTED_MODEL_BEHAVIOR;\n\n      // Skip if caching is disabled\n      if (!enableCaching || !request.model) {\n        return handler(request);\n      }\n\n      const isAnthropicModel =\n        request.model.getName() === \"ChatAnthropic\" ||\n        (request.model.getName() === \"ConfigurableModel\" &&\n          (request.model as ConfigurableModel)._defaultConfig?.modelProvider ===\n            \"anthropic\");\n      if (!isAnthropicModel) {\n        // Get model name for better error context\n        const modelName = request.model.getName();\n        const modelInfo =\n          request.model.getName() === \"ConfigurableModel\"\n            ? `${modelName} (${\n                (request.model as ConfigurableModel)._defaultConfig\n                  ?.modelProvider\n              })`\n            : modelName;\n\n        const baseMessage = `Unsupported model '${modelInfo}'. Prompt caching requires an Anthropic model`;\n\n        if (unsupportedModelBehavior === \"raise\") {\n          throw new PromptCachingMiddlewareError(\n            `${baseMessage} (e.g., 'anthropic:claude-4-0-sonnet').`\n          );\n        } else if (unsupportedModelBehavior === \"warn\") {\n          console.warn(\n            `PromptCachingMiddleware: Skipping caching for ${modelName}. Consider switching to an Anthropic model for caching benefits.`\n          );\n        }\n        return handler(request);\n      }\n\n      const messagesCount =\n        request.state.messages.length + (request.systemPrompt ? 1 : 0);\n\n      if (messagesCount < minMessagesToCache) {\n        return handler(request);\n      }\n\n      /**\n       * Add cache_control to the last message\n       */\n      const lastMessage = request.messages.at(-1);\n      if (!lastMessage) {\n        return handler(request);\n      }\n\n      const NewMessageConstructor =\n        Object.getPrototypeOf(lastMessage).constructor;\n      if (Array.isArray(lastMessage.content)) {\n        const newMessage = new NewMessageConstructor({\n          ...lastMessage,\n          content: [\n            ...lastMessage.content.slice(0, -1),\n            {\n              ...lastMessage.content.at(-1),\n              cache_control: {\n                type: \"ephemeral\",\n                ttl,\n              },\n            } as ContentBlock,\n          ],\n        });\n        return handler({\n          ...request,\n          messages: [...request.messages.slice(0, -1), newMessage],\n        });\n      } else if (typeof lastMessage.content === \"string\") {\n        const newMessage = new NewMessageConstructor({\n          ...lastMessage,\n          content: [\n            {\n              type: \"text\",\n              text: lastMessage.content,\n              cache_control: {\n                type: \"ephemeral\",\n                ttl,\n              },\n            },\n          ],\n        });\n        return handler({\n          ...request,\n          messages: [...request.messages.slice(0, -1), newMessage],\n        });\n      }\n\n      throw new PromptCachingMiddlewareError(\n        \"Last message content is not a string or array\"\n      );\n    },\n  });\n}\n"],"mappings":";;;;AAOA,MAAM,yBAAyB;AAC/B,MAAM,cAAc;AACpB,MAAM,gCAAgC;AACtC,MAAM,qCAAqC;AAE3C,MAAM,gBAAgB,EAAE,OAAO;CAK7B,eAAe,EAAE,SAAS,CAAC,UAAU;CAKrC,KAAK,EAAE,KAAK,CAAC,MAAM,IAAK,EAAC,CAAC,UAAU;CAKpC,oBAAoB,EAAE,QAAQ,CAAC,UAAU;CAQzC,0BAA0B,EAAE,KAAK;EAAC;EAAU;EAAQ;CAAQ,EAAC,CAAC,UAAU;AACzE,EAAC;AAKF,IAAM,+BAAN,cAA2C,MAAM;CAC/C,YAAYA,SAAiB;EAC3B,MAAM,QAAQ;EACd,KAAK,OAAO;CACb;AACF;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwHD,SAAgB,iCACdC,mBACA;AACA,QAAO,iBAAiB;EACtB,MAAM;EACN;EACA,eAAe,CAAC,SAAS,YAAY;;;;GAInC,MAAM,gBACJ,QAAQ,QAAQ,QAAQ,iBACxB,mBAAmB,iBACnB;GACF,MAAM,MACJ,QAAQ,QAAQ,QAAQ,OAAO,mBAAmB,OAAO;GAC3D,MAAM,qBACJ,QAAQ,QAAQ,QAAQ,sBACxB,mBAAmB,sBACnB;GACF,MAAM,2BACJ,QAAQ,QAAQ,QAAQ,4BACxB,mBAAmB,4BACnB;AAGF,OAAI,CAAC,iBAAiB,CAAC,QAAQ,MAC7B,QAAO,QAAQ,QAAQ;GAGzB,MAAM,mBACJ,QAAQ,MAAM,SAAS,KAAK,mBAC3B,QAAQ,MAAM,SAAS,KAAK,uBAC1B,QAAQ,MAA4B,gBAAgB,kBACnD;AACN,OAAI,CAAC,kBAAkB;IAErB,MAAM,YAAY,QAAQ,MAAM,SAAS;IACzC,MAAM,YACJ,QAAQ,MAAM,SAAS,KAAK,sBACxB,GAAG,UAAU,EAAE,EACZ,QAAQ,MAA4B,gBACjC,cACL,CAAC,CAAC,GACH;IAEN,MAAM,cAAc,CAAC,mBAAmB,EAAE,UAAU,6CAA6C,CAAC;AAElG,QAAI,6BAA6B,QAC/B,OAAM,IAAI,6BACR,GAAG,YAAY,uCAAuC,CAAC;aAEhD,6BAA6B,QACtC,QAAQ,KACN,CAAC,8CAA8C,EAAE,UAAU,gEAAgE,CAAC,CAC7H;AAEH,WAAO,QAAQ,QAAQ;GACxB;GAED,MAAM,gBACJ,QAAQ,MAAM,SAAS,UAAU,QAAQ,eAAe,IAAI;AAE9D,OAAI,gBAAgB,mBAClB,QAAO,QAAQ,QAAQ;;;;GAMzB,MAAM,cAAc,QAAQ,SAAS,GAAG,GAAG;AAC3C,OAAI,CAAC,YACH,QAAO,QAAQ,QAAQ;GAGzB,MAAM,wBACJ,OAAO,eAAe,YAAY,CAAC;AACrC,OAAI,MAAM,QAAQ,YAAY,QAAQ,EAAE;IACtC,MAAM,aAAa,IAAI,sBAAsB;KAC3C,GAAG;KACH,SAAS,CACP,GAAG,YAAY,QAAQ,MAAM,GAAG,GAAG,EACnC;MACE,GAAG,YAAY,QAAQ,GAAG,GAAG;MAC7B,eAAe;OACb,MAAM;OACN;MACD;KACF,CACF;IACF;AACD,WAAO,QAAQ;KACb,GAAG;KACH,UAAU,CAAC,GAAG,QAAQ,SAAS,MAAM,GAAG,GAAG,EAAE,UAAW;IACzD,EAAC;GACH,WAAU,OAAO,YAAY,YAAY,UAAU;IAClD,MAAM,aAAa,IAAI,sBAAsB;KAC3C,GAAG;KACH,SAAS,CACP;MACE,MAAM;MACN,MAAM,YAAY;MAClB,eAAe;OACb,MAAM;OACN;MACD;KACF,CACF;IACF;AACD,WAAO,QAAQ;KACb,GAAG;KACH,UAAU,CAAC,GAAG,QAAQ,SAAS,MAAM,GAAG,GAAG,EAAE,UAAW;IACzD,EAAC;GACH;AAED,SAAM,IAAI,6BACR;EAEH;CACF,EAAC;AACH"}